A number of new models have been made available in the platform over the last month:

• Meta's Muse Spark 1.1
• Thinking Machines's Inkling
• DeepSeek's V4 Flash

SCIM Provisioning​

A major update to SCIM has been added to support creating, updating, de-activating and removing Abacus users automatically, as well as syncing directory groups. A new option in the dev platforms settings is available for admins to configure this:

For more information please read our documentation

Voice Agents​

Voice Agents are AI assistants that you can interact with using real-time speech instead of typing. They combine advanced speech-to-text, reasoning, and text-to-speech technologies to enable fluid, human-like verbal conversations. On the Abacus AI platform, these agents typically offer:

• Real-Time Verbal Interaction: You can speak directly to the agent (using the microphone icon in the chatbot or a dedicated mobile interface) and hear a spoken response with very low latency.

• Dedicated Phone Line: Provision a dedicated phone number to the agent to allow users to call in directly.

• Multimodal Capabilities: The agent can "listen" to your instructions, process complex data or code, and explain the results verbally while simultaneously showing relevant text or artifacts on your screen.

• Mobile Accessibility: They are particularly useful on the ChatLLM mobile app, allowing for hands-free productivity while you are on the go.

Enabling agents for voice interaction is done as part of a custom bot. This can be configured manually using Enable Voice option in the bot configuration screen or via the Abacus AI Agent:

To provision a phone number to the agent requires additional setup steps. Once deployed navigate to the Super Assistant UI to interact with the agent, where a new Add phone number option will be available:

This will open a new window in the Bot settings page for creating and managing the phone number for this agent:

Phone numbers can be managed on an agent-by-agent basis in the settings page:

Custom Bot Settings​

As part of the UI settings for custom bots we have added two new options for controlling access to the Humanize and Verification tools. Navigate to the bot settings menu found under Bots in the Super Assistant settings:

Under Open Settings there are two new toggles for managing these tools:

Self-Improving Tasks​

An exciting new feature for the AI Agent is the ability to create self-improving tasks that recursively work to improve the outcome. While normal tasks handle recurring work, self-improving tasks work to get measurably better at a task over time. Each run makes one focused change, an independent evaluation measures the result and the next run builds on what actually worked.

To create a self-improving task use the menu option on the left of the Super Assistant homescreen, and then click on the Self-improving AI banner above the prompt window:

We provide some examples to get you started or you can create your own. What’s important when creating these tasks is to be clear in the prompt how the task should execute:

• What to improve (the objective)
• What the agent may change (e.g. code, prompts, data, replies…)
• How to measure it (the metric and where the data comes from)
• How often to run
• Any limits or approvals (e.g. “open a PR, don’t merge”, “max one post per day”)

In this example we prompted the agent task to score incoming lead data from a CRM. Over multiple runs assessing conversation data and new leads it improves its overall precision in identifying the most promising customers:

For more details please reference the documentation

Custom Router Updates​

Several updates have made to the new custom router functionality:

Firstly, custom chatbots can now be added as one of the options that a prompt gets routed to. Within the LLM selection for a particular router classification a new tab is available that lists all deployed bots:

This option provides extended flexibility to ensure users are re-directed to the correct source, without them having to select chatbots independently. It also more seamlessly marries the underlying power of the Super Assistant with any specific solutions you have designed.

In addition custom router access can now be permission controlled through access groups in the same way as custom bots, giving added flexibility to design routers for specific groups within your organization.

To manage this navigate to the Profile >> ORGANIZATION >> Bots menu in Super Assistant:

Use the group icon column to manage access permissions for each custom router.

User-Level Connectors Management​

Connector configurations are used to manage the default scopes and settings of the connections end-users are authenticating in the Super Assistant. These had previously been part of the developer platform but have now been added to the User Connectors menu in the Super Assistant.

To create a connector configuration navigate to the Connectors page through your profile menu:

Once on the User Connectors tab a Connector Configurations option will be available below the list of user connectors that have already been authenticated. Click on Add Config to view a list of the available connectors to configure (n.b. Not all user level connectors have an associated configuration connector available).

Select a connector to configure. Depending on the application the options will differ, but for Outlook the configuration options will look like this:

Click Save to create that configuration. To apply this configuration to all users you will need to Connect:

Any users connecting to that service through the Super Assistant will now be restricted to the configuration established in the connector configuration. Existing connections authenticated will be asked to re-authenticate in order to apply these changes.

This feature allows platform admins to control which connectors are available for end-users to authenticate to. It is found under the Settings and Permissions >> Features & Integrations menu:

Admins can use the toggles to manage how connectors are available in the organization, with three STATUS levels:

• Enabled - the connector is available at both a user and organizational level
• User-level off - the connector is only available as an organizational level connector in the developer platform
• Disabled - the connector is not available

The other new related feature is also found on the same Features & Integrations page. In order to enforce authentication of connectors with only approved organizational domains admins can turn on the Enforce Connector Email Domain option:

Domains are configured as part of your organization setup - for more details on managing the domains please contact your account's Forward Deployed Engineer.

Logs Export​

Abacus now provides the ability to directly export usage and admin logs. Under Settings & Permissions >> Audit Logs admins can configure how the logs should be exported:

Standard options include:

• Abacus managed storage
• Cloud storage bucket (customer owned)
• Email
• Splunk

Once configured the export runs every hour for the previous hour’s activity.

Global Context​

In order for organizations to be able to provide standard prompts for all interactions in the Super Assistant a new Global Context option has been added. To provide global context navigate to the Settings and Permissions >> Features & Integrations menu in the Super Assistant.

Note that Global Context is injected into every conversation and will increase token usage on every message. The context may conflict with existing system prompts or a user’s own custom instructions causing unexpected behaviour.

Any prompts here should be kept short and focused.

App Tools​

Several tools have been added as part of the Application building process with the Abacus Agent. Once the application is in preview mode a Tools menu is available with the ability to run:

• Test Agent - The Test agent runs automated QA tests on your app and can set up a testing scheduled task that re-runs them regularly
• SEO Agent - The SEO agent audits your deployed app and helps optimize it for search engines. It will review your Lighthouse SEO score, meta tags and descriptions, sitemap, and ease of crawling, then apply improvements to your app.
• Self-Healing Agent - Scan production logs daily for backend errors and email you a digest. Fixes are staged on a separate URL so production is never touched

Coming Soon...​

In order to provide more flexibilty to manage work on the platform admins will be able to transfer ownership of individual projects. Previously all projects owned by a particular user had to be transferred, with no ability to sub-select particular ones.

To transfer projects use the View ChatLLM Projects tab, found under Settings & Permissions. A new Transfer tab is available:

A number of new models have been made available in the platform over the last month:

• Anthropic's Sonnet 5
• OpenAI's GPT5.6 (Sol, Luna, Terra)
• Grok 4.5

Livebench​

We have launched a new version of our LiveBench site for LLM assessment. Here are the main updates:

Price per task A new column in the leaderboard tracks the overall cost per task for each model, providing you with much clear information for cost optimization when selecting an LLM.

The costs get filtered for the specific categories you are looking at. So looking at Agentic Coding allows me to sort models based on each subcategory, but also see the costs of these models on this category of tasks:

Category details It is now much easier to view how each model performs on the category sub-tasks by simply clicking on the model to view this information:

Build Your Own Router​

An exciting new feature in the platform is the ability to create your own custom routers. Access to this feature is found under Profile >> Custom Routers and APIs menu:

To create your own custom router click on the New Custom Router button. This will give you three main options:

1. Describe with AI Use the Abacus Agent to build your routing mechanism through natural language. For example take this prompt:

The Agent digests this prompt and creates the following configuration for the router, defining the categories and creating category descriptions that are provided as classification prompts:

Notice that categories can be further refined, deleted or new ones added by the user.

2. Use a template For this option Abacus provides a number of templates for users to get started with pre-defined categories, descriptions and models:

3. Start from scratch This option gives users a totally blank slate to fully design their own router from scratch.

Once your custom router is built it will be available in two places: via the routellm api and as part of the Super Assistant chat mode.

Via the RouteLLM API When the router is created it is given an api tag:

This tag can be plugged into the routeLLM api to identify the model that is being used in the api request.

Details about how to use the routeLLM api can be found either on the Custom Routers and API page or in our documentation

In the Super Assistant UI The custom router will be available as one of the models available in chat mode:

Organizational MCP Servers​

A key addition to our MCP server options is the ability to now make them organization wide. Platform administrators have the option when configuring a connection to an MCP server to make it available to all users in the organization.

Viewing Organizational MCP Servers First navigate to the MCP Server Configuration tab, found under Profile >> Connectors. On this page you will be able to see MCP servers that you have personally configured, plus a list of organization MCP configurations:

An important point to note is that these organizational MCP servers still require individual users to Connect to the server and provide authentication if necessary.

Creating an Organizational MCP Configuration Platform admins can create these servers using the Add Server button on the MCP Server Configuration page.

Select either the Add via URL or Add via JSON option. Complete the details based on the option and select the Make available to the whole organization option at the bottom of the configuration box:

For more detailed information on this feature please refer to our documentation

App Ownership Transfer​

A new App Ownership Transfer feature makes it easier to manage applications as teams grow and responsibilities change. App owners can now transfer ownership of both deployed and hosted apps to another user within the same organization, without affecting the app's deployment or permissions. This helps organizations maintain continuity when team members change roles, leave the organization or hand over responsibility for an application.

When a transfer is completed:

• The selected user becomes the new owner of the application
• The ownership change is reflected throughout the platform wherever the app owner is displayed
• Existing app permissions remain unchanged

To transfer ownership of an application first navigate to the Apps Access tab found under Profile >> Permissions and Groups. Find the relevant application and click on the settings icon:

On the settings menu use the Transfer ownership button:

Microsoft Office Add-ins​

The power of Abacus Agent is now available directly inside your main Microsoft Office applications as an Add-in

Some of the key features:

• Excel: Read multi-tab workbooks to help you build financial models and data sheets. It updates assumptions without breaking your existing formulas and precisely tracks every cell it modifies.

• Word: Draft new documents, summarize lengthy reports, and redline contracts. AI edits land natively as tracked changes, and the assistant can even respond to comment threads to explain exactly what it changed and why.

• PowerPoint: Generate images, create native, fully editable charts, and restructure entire decks. The AI works inside your specific template—respecting your slide master layouts, fonts, and colors.

• Microsoft Teams & Outlook: Summarize long email threads or chat conversations, draft context-aware replies, and collaborate faster across your entire team.

PDF Analysis Tool​

We have made a significant update on how the platform conducts analysis of PDFs, allowing users to complete complex tasks with simpler models that otherwise was not possible.

The platorm now reads and reasons over an uploaded PDF end-to-end. Ask a question about a document - or just say "summarize this" - and it returns a synthesized answer grounded in the file's contents. It works on documents that chat models normally can't handle well, including scanned/image-only PDFs, and it reads the whole document rather than just the first few pages, so it reliably picks up fine print, figures, and specific values (e.g. amounts, dates, clauses, table cells).

Some of the key features:

• Works on scanned and native PDFs, including long documents (no page limit)
• Answers whole-document questions and summaries, no just keyword lookups
• Handles multiple uploaded files and specific page ranges
• Automatically kicks in when a PDF is attached

Admin Settings Updates​

A range of new admin controls have been added to the platform to help organizations manage usage limts, authentication options and monitoring.

Monthly Usage Limits A new option has been added to the settings for usage limits that allows organizations to choose between setting daily and monthly limits. This can be found under the Profile >> Permissions and Groups >> Settings and Permissions >> Data & Usage menu:

User Connectors - Admin Console To increase visibility of which First Party Connectors are being used a new admin tab has been added.

Under Profile >> Permissions and Groups >> Settings and Permissions a new tab called View First Party Connectors is available. This provides information on the connector type, person who authenticated it, current status and created date:

User Connectors - Email Domain Restriction A new security control, Enforce Connector Email Domain, has been added to give administrators granular control over which accounts can be used to authenticate external integrations. This feature ensures that users only connect professional or approved accounts to the platform, preventing the accidental use of personal email addresses or unapproved third-party domains when setting up connectors.

Access to this feature is found in Permissions & Groups > Setting & Permissions > Enforce Connector Email Domain:

Agent Firewall This new enterprise security feature gives organizations greater control over the network access available to the Abacus AI Agent during task execution. Rather than allowing unrestricted outbound internet access, administrators can define firewall rules that limit which destinations Agent VMs are permitted to access. This enables organizations to align agent execution with their existing security and compliance requirements.

Agent Firewall settings are available under the Profile >> Permissions and Groups >> Agent Firewall menu:

The Abacus team will be hosting a webinar on Wednesday June 24th, 9am PST, focused on using the platform to build AI Workflows and custom Chatbots. You can register for the event here.

Developer Platform

Connectors​

Microsoft Purview Labels The platform now supports using data labels defined in Purview to manage which data should be processed by the Abacus platform. When authenitcating Sharepoint or OneDrive a new parameter, Skip Sensitivity Labels, allows you to list all document labels that should not be processed:

Newly Available We have added a number of new connectors to the platform that are available either on the developer platform or as First Party Connectors in the Super Assistant. The following connectors are all available to use:

• Dropbox
• HubSpot
• Athena
• QuickBooks
• Tableau
• DocuSign

For more information on any of these please refer to our documentation page

Super Assistant

New Models​

A number of new models have been made available in the platform over the last month:

• Anthropic's Claude Opus 4.8
• Google's Gemini 3.5 Flash
• Qwen 3.7 Max

To find out how a range of models perform don’t forget to check out our benchmark, LiveBench

Abacus Design​

Design has been updated to provide a much more flexible experience during the design phase of projects, as well as more options of what to do with your completed work.

Edit Mode For more flexibilty to manually edit designs built by the Abacus Agent users can now use the Edit mode to make a range of colour, text and formatting changes directly on the design.

Once the Agent has completed the design and you are able to preview the results use the Edit button at the top of the preview pane to enable inline editing:

• The editing pane will appear on the right-hand side of the screen, giving you options to change the content, topography, layout and colours of the elements that you select in your design
• To further refine which components you are editing use the Layers option at the top-left of the preview pane. This allows you to select the specific html components you wish to edit

Export Options Once your designs are complete there are a number of options for you to be able to export your completed design:

Note that if you wish to work on a design already created in Figma you can provide the Abacus Agent with the relevant Figma url and it will be able to re-create the work inside of Abacus. Once your work in Abacus is complete you can use the SVG option to export the designs back to Figma

Turning Designs into Apps When you decide that your design is ready to be made into a fully functioning application this can be done using the Abacus Agent. We have added a new Build button to allow you to do this easily - it's available in the conversation whenever you have a version of your design to preview:

Abacus Apps​

A new feature for Abacus Apps that adds significant benefits when building internal facing applications is the passing through of user information via headers. Apps can be designed to grab information from the headers about the specific user interacting with the application. At present this includes:

• The user email associated with their Abacus account
• Any tokens from First Party Connectors they have already authenticated

This makes it very easy to build applications integrated with common user tools without needing to embed complex authentication methods or have users re-authenticate into these applications. To work it is required that the app be deployed as either Owner Only or Private to Org.

Here is simple prompt that demonstrates how this feature can be used:

Build me an app that allows me to view all my latest activity in Slack and Gmail as well as show all of my Google Calendar events for the day.  The app should:

1. Create different sections in one dashboard page for each of these applications
2. Show me the last 10 messages in Slack and last 10 emails from Gmail
3. Show an error message for users that have not authenticated these apps already.  The message should redirect them to the connectors page in the platform.

There should be no additional SSO built into the app and users should not be asked to login. Credentials should be directly imported from their existing connectors if available.  
The tokens and user email are available as part of the headers when they access the application and can be used to access data from the connectors.

Abacus CoWork​

MCP Integration Abacus CoWork now supports MCP integration in both CoWork and Code modes. To configure servers use the MCP Servers option in the settings toolbar:

There are two options for adding MCP servers into the tool: 1. Add Server allows you to add a new MCP server configuration. This can either be STDIO for local servers or Streamable HTTP for remote servers:

2. Importing an existing server configuration file is also supported. This can be done as a manual upload or imported from Cursor, Claude Code or the Abacus CLI

For more details please refer to our MCP documentation here. Note also that Browser Use has its own settings options found in the same toolbar.

Custom Chatbots Custom chatbots that you have access to in the web application are now also available in Chat mode. To access these bots use the model dropdown menu in the prompt window:

Below the list of regular LLMs will be all the custom bots that a user has access to. A new tab for filtering to only these bots is also present:

Global Skills​

Custom skills that have either been built in Abacus or imported can now be made available to the entire organization as a global skill. To make a skill global a platform admin will need to use a new option in the Super Assistant permissions page: Profile >> Permissions and Groups >> Permissions. Under the Features & Integrations tab is a new option for managing global skills:

Note that global skills will initially appear as always-enabled for every user in the organization. Each global skill increases token usage for every conversation and skill instructions may conflict with existing system prompts or other skills, potentially causing unexpected behavior.

Users will be able to see global skills that have been added in the UI using the small spanner symbol in the prompt box. They can toggle the global skills on/off on a per conversation basis should they wish to.

Admin Updates​

As well as the global skills several new admin controls have been added to the platform this month. All of these options can be viewed in the Permissions tab, accessed through the Profile menu >> Groups and Permissions >> Permissions:

• Manage Skill Creation allows admins to limit which users are able to build or import skills. Combined with global skills this create an effective way for the organization to manage which skills are created and available for users. This option is found under the Access & Restrictions tab

• Email Domain Restrictions gives the option to restrict which email domains the built-in Abacus mailer tool can send emails to, which provides a more secure environment for task execution. This option is found under the Access & Restrictions tab

Note that this domain restriction only applies to the Abacus mailer tool and does not apply to any first party connector tools like Gmail or Outlook.

• Manage First Party Connectors is a simple interface to manage which first party connector tools are available for the business to authenticate. It is found under the Access & Restrictions tab:

• Global Project Sharing provides additional capabilities for managing how projects created by users can be shared. When turned on this option limits users from sharing projects globally, i.e. with the entire organization - they will only be able to share projects with specific groups or individuals. The settings are found under the Sharing & Privacy tab

User Spend Limit Monitoring​

Users can now monitor their use against any daily limits that have been put in place. Users can click on their profile picture and then under the Profile menu will be able to see theor current daily spend:

Note that this will only be visible if user limits have been put in place via the Rate Limts setting under Permissions

The Abacus team will be hosting a webinar on Wednesday May 20th, 9am PST, focused on using the platform to build AI Workflows and custom Chatbots. You can register for the event here.

Super Assistant

April brought some major expansion of Abacus AI Agent capabilities — including multi-agent collaboration with Agent Swarms, smarter cost controls with Effort Levels, richer app-building tools with Design and conversation-level Skills selection. As well as introducing a new app only role.

New Models​

A whole host of new models have been made available in the platform over the last month:

• Anthropic's Claude Opus 4.7
• OpenAI's GPT-5.5 (including the Thinking, Pro and Instant versions)
• X.AI's Grok 4.3
• Kimi K2.6 Thinking
• Deepseek V4 Pro & V4 Flash

To find out how a range of models perform don’t forget to check out our benchmark, LiveBench

Abacus Design​

A new feature for optimization of building UI designs and wireframes is Abacus Design, which is available as part of the Abacus Agent. Instead of using the Agent to build fully fledged apps it can be prompted instead to build only the design elements of an app or dashboard. This significantly reduces token use for projects where you want to iterate on a design or built multiple variations, before transitioning to a full application.

Effort Levels​

You now have direct control over how much effort is used during Super Assistant chats and AI Agent tasks. The Effort Level option is available when using RouteLLM in chat mode, or when you select Agent mode, and gives you the option between Auto, Low Effort and High Effort. Low Effort mode routes to using cheaper models (with lower thinking levels, when required), which allows you to optimize the cost of completing simpler tasks that don't require more expensive compute.

Keeping this set to Auto is the default recommendation for balancing cost and performance effectively.

Agent Swarms​

Abacus AI Agent can now orchestrate multi-agent swarms — coordinating several AI Agents to tackle multiple tasks collaboratively. Take this prompt as an example:

Build a full-stack personal finance ecosystem consisting of: A mobile app (iOS + Android) for daily expense tracking with features like smart categorization, receipt scanning, recurring expense detection, and real-time insights. A web app that acts as a power-user dashboard with advanced capabilities such as: Interactive visualizations, AI-powered spending analysis and anomaly detection and Budget planning. Ensure both apps are seamlessly synced via a shared backend.

With normal Abaucs Agent use this would require building one application first and then using the database to seed a new conversation to build the second application. This approach is much less efficient than the swarm approach as it requires you to build the apps in succession, rather than in parallel. But more importanly swarms are able to assess requirements for both applications simultaneously when designing and building the shared infrastructure, which means fewer iterations and re-builds than before.

Note that although there is a low-effort swarm mode available this can be a very token intensive process, so should be used with caution.

To activate swarm mode the task must be started from inside a project. Aabcus will then create multiple conversations for each ot the subtasks, with one master conversation managing the entire process:

Conversation-Level Skills​

This new option allows you to select specific skills the Abacus Agent should use for a particular task, rather than relying on it selecting from the global set of skills available. Skills pre-selection like this removes the non-determinism of the Agent selecting the correct skill, especially where you have multiple similar skills available for it to use.

To enable this use the small wrench button when selecting Agent mode.

App-Only Role​

A new App Only user role is now available for AppLLM. This allows administrators to grant users access exclusively to deployed apps — without exposing the broader ChatLLM workspace, agent capabilities, or admin settings. Perfect for distributing internal tools or customer-facing apps to users who only need to interact with the finished product.

This role can be manually assigned in the Team menu: navigate to Profile >> Groups and Permissions >> Team. For customers using SSO a new app_user role can be created, see our documentation for more details.

Admin Updates​

Several new admin controls have been added to the platform this month. All of these options can be viewed in the Permissions tab, accessed through the Profile menu >> Groups and Permissions >> Permissions:

• View Shared Projects Admins are now able to view all projects in the organization that have been shared with specific users or groups. Use the View ChatLLM Projects tab to see information about the projects themselves and which users they have been shared with

• Manage User Tasks The View User Tasks tab allows admins to now see and manage all tasks that are running in the organization

• Manage Media Generation Access This option allows admins to control who can access image, video and speech models for generating content

When connecting to Azure Storage, service principal has been added as a new authentication method. To configure Azure Storage navigate to the Connected Services tab, accessed via the Profile menu >> Manage Connectors

Super Assistant

New Models​

A number of new models have been made available in the platform over the last month: Google Gemma 31B, GLM 5.1, MiniMax M2.7 and Qwen 3.6. To find out how a range of new models perform don’t forget to check out our benchmark, LiveBench

Abacus AI Agent​

Abacus AI Agent is the new name for Deep Agent. The name is being updated across the platform and will be fully rolled-out in the coming weeks.

App Deployment and Task Controls Two new admin controls have been added to enhance controls over which users in an organization can deploy applications and create automated tasks.

To enable these options use the Permissions menu in the Super Assistant UI, accessed via: Profile >> Groups and Permissions >> Permissions

Manage Agent Video Exmaples For organizations that would like to remove the video exmaples from the AI Agent workspace, a new option is available under the Permissions menu:

Deep Research Verification When the Abacus Agent completes a research task a new option appears in the preview screen allowing users to perform a fact verification check on the content.

The agent will perform the verification analysis and provide a report to the user on the quality of the research, covering metrics such as: statement classification into supported, unsupported, unrelated and contradicted; citation accuracy; source relevance and factual consistency.

These metrics are provided as both high level statistics,

as well as more detailed comparisons for specific statements:

Once the verification has been completed the preview panel will show a Fact Checked label:

Several updates have been made to powerpoint generation in the Abacus Agent covering editing, speaker notes and direct publishing options.

Slide Editing The editing function for slides has been enhanced to provide considerably more options for users handling in-line edits in the preview screen. Use the Edit button at the top of the screen to open the editing interface, whereby the editing toolbar will appear on the right of the screen:

Speaker Notes Prompt the Abacus Agent to add speaker notes if you would like the additional context to accompany each slide. Notes are added under the slide images in the preview panel:

Slide Publishing A new option has been added to enable slide publishing for easier sharing and presentation. Once the slides are complete use the Publish button at the top of the screen:

Once published the presentation is available at a dedicated url and can be viewed as a slideshow directly from that location:

Skills​

A significant new feature is the introduction of skills to the Abacus platform. Skills give the Abacus AI Agent the ability to execute with specific instructions or domain expertise in a consistent manner. They ensure that there is:

• Consistent Quality - The agent follows established methods every time
• Specialized Expertise - You can provide specific instructions on how it handles technical writing, data analysis, code review and more
• Automatic Activation - The Abacus AI Agent detects when to use each skill based on your request
• Time saving - Skills streamline the prompting process avoiding repeatedly needing to define prompts

Skills can be simply defined as a SKILL.md file, but can also be configured with separate folders that contain examples, templates and references. More details on how to create skills are available on our documentation page.

Managing Skills Adding and managing skills in the platform can be done through the Customize & Add Skills menu, which is found when clicking on your profile icon in the top-right of the screen:

Then select the Agent tab:

Under the Skills section you will see any custom skills that you have already added as well as any Abacus built system skills that you have imported. For custom skills a toggle is available to turn them on or off, but system skills you have imported are on by default - to turn these off delete that skill, it can be re-imported later if required.

Importing System Skills To import Abacus built system skills click on the Import Skill button:

This will open a new window listing all of the available system skills to import. Once imported they will be automatically available for the Abacus AI Agent.

Adding Custom Skills If you have defined your own custom skills, these are added using the New Skill button:

New skills can be added either by directly uploading a .zip file or by providing a valid GitHub repo url. For both options the SKILL.md file must be at hte root level.

Custom skills can be managed in the main menu, allowing you to:

• Edit existing skills by uploading new versions of the skill files
• Delete the file completely
• Toggle custom skills on or off

Both system and custom skills can also be added to projects using the Customize & Add Skills option:

Note: currently skills are only available for use by the Abacus AI Agent for standalone tasks, or AI Agent tasks inside of projects. Skills added to projects won't be used for regular, chat mode, conversations

Abacus Desktop - Cowork​

Abacus AI CoWork brings agentic, autonomous task execution to your desktop — extending Abacus AI Agent's power beyond coding into general knowledge work. CoWork can plan, break down, and execute complex, multi-step tasks directly on your local files without requiring terminal access or manual uploads.

CoWork has the following key capabilities:

• Direct Local File Access: Read, write, organize, and transform files on your machine — documents, spreadsheets, presentations, images, data files, and more.
• Sub-Agent Coordination: Complex tasks are automatically decomposed into subtasks and executed in parallel for faster completion.
• Professional Outputs: Generate polished, fully editable deliverables — Excel spreadsheets with working formulas, formatted Word documents, PowerPoint presentations, PDF reports, and more.
• Long-Running & Scheduled Tasks: Handle tasks that require extended execution without conversation timeouts. Schedule recurring tasks to run automatically on a chosen cadence.
• Browser Integration: Combine local file operations with web-based workflows — fill forms, navigate sites, and pull data from online sources seamlessly

CoWork can be accessed as part the Abacus AI Desktop suite. Us the download link at the bottom of the Super Assistant screen:

CoWork provides two modes for users to interact with: Chat and CoWork. Chat mode provides a standard conversational mode for users to leverage either RouteLLM or a number of specific LLMs. For more complex tasks that leverage the Abacus AI Agent select CoWork mode - here users can decide how much effort the system should use to solve the task, with options of Low, Medium and High

For additional information about Abacus Desktop please see our documentation pages.

Project Sharing​

An additional option has been added within the project sharing options to make access read-only:

In order to support more complex RAG applications that require page level context, or where you have complex PDF's that contain images, tables and text all on one page, new options have been added during the indexing and configuration steps for custom chatbots.

To enable this capability, it must first be toggled on when building the Document Retriever

Options such as chunk size are not relevant as each page is treated as a single chunk.

After adding page level Document Retrievers to a chatbot configuration additional options are available in the Unstructured Data Options

• Max Search Results - This handles how many page results are returned from the search - based on the top scoring pages
• Score Cutoff - Filters out search results below this threshold. This is applied before Max Search Results
• Max Page Images - Sets the number of page images that are also sent to the LLM, alongside the page chunk. For complex PDF documents with multiple content types on the same page this option provides both the extracted text and image to the LLM for improved understanding.

N.B. Max Page Images only works for PDF files - text or image files will not got passed as images to the LLM

Custom Tool Creation​

When building custom tools during AI Workflow creation users are now redirected to Abacus DeepAgent for a more complete vibe-coding experience.

Adding custom tools can be done in the workflow editor. Use the Add Node menu and select Tool Node:

In the tools menu, scroll to the bottom and select Create new tool:

A new window will appear inside the workflow editor where the DeepAgent will prompt the user to describe the requirements for the tool, including inputs, outputs and any connectors needed. Once the user has sepcified the tool DeepAgent will build the custom tool:

Users can test the tool, make modifications to iterate until the desired functionality is achieved and ultimately register the tool all within the same conversation.

Slack Config Connector​

A new config option has been added for our Slack connector. This allows you to specify the scopes that the end-user Slack connector tools will have available. To enable this option navigate to the Connected Services tab, accessed via the Profile menu >> Manage Connectors

Once configured users authenticating into Slack via the tool in Super Assistant, or using the tools deployed via chatbots or workflows will only have access to the scopes specified in the configuration.

New Models​

OpenAI's GPT-5.4 and GPT-5.3-Codex have been added to the platform. To find out how a range of new models perform don’t forget to check out our benchmark, LiveBench

Company Knowledge​

Company Knowledge has been added as a new chat mode for end users in the Super Assistant. It's designed to improve the performance of searching across internal data by allowing users to select which applications to search against for each prompt. This improves the accuracy of the search and reduces latency caused by searching across unneccessary applications.

Enabling Company Knowledge is straightforward - select it from the chat mode drop-down at the bottom of the prompt window:

Once selected users can choose which applications to search across for each prompt:

Note that within the same conversation users can switch between Company Knowledge and regular Chat Mode seamlessly.

DeepAgent - Production Databases​

An additional option now available with app development is to create a production database to allow users to safely iterate on database changes in dev mode:

This feature provides a number of key benefits:

• Easily promote dev databases to production
• Isolation of production data to help prevent accidental data loss during development and testing
• Any changes that might lead to data loss require additional approvals
• Ability to sync data from production to the dev database

DeepAgent - HTML Presentations​

We have updated how DeepAgent creates presentations to allow much more flexibilty during the building and editing process.

By using HTML to build the final presentation our existing Edit function can be leveraged to edit the presentation slides inline more efficiently. The colours and fonts used to style the slides is also defined clearly during the build process, so can be much more easily specified by users:

DeepAgent Excel Tool​

Another DeepAgent tool that has been updated is the Excel tool, which now allows users to perform more advanced tasks directly inside the uploaded spreadsheet.

Simply upload your spreadsheet to DeepAgent and prompt the changes you would like to be applied: from adding formulas and calculations to more detailed analysis and visualizations.

DeepAgent provides the completed spreadsheet back to users for downloading.

We will be hosting a webinar to showcase the latest developments with DeepAgent on February 25th 9am - 10am PST. If you would like to attend please register on our associated Eventbrite page

AI Workflow Updates​

A number of updates have been made to the AI Workflow builder to simplify the creation process.

A new feature of the workflows are the LLM Agent Nodes which incorporate the power of the Abacus platform to handle common agent tasks like tool integration and data processing without having to build them each time. In the workflow builder select Add Node and then LLM Agent Node from the options:

The system will give you two options for adding the node into the workflow: New or Existing. When adding a New Agent Node select the LLM from the dropdown list:

If you have an existing Agent Node you would like to add, or you have a custom ChatBot you have previously deployed that you would like to incorporate, selecting Existing will allow you to select the specific project and deployment:

Click Generate to add the new node into the workflow. To update any of the configuration for that node, make sure that node is selected and click on Node Configuration in the top-right of the screen. This will provide a number of key options:

• LLM Options -The ability to easily change the LLM you would like to use
• Add Tools - Add any of the pre-defined connector tools
• Behaviour Instructions - Provide prompt instructions to guide the LLM's behaviour
• Knowledge Base - Attach any specific document retrievers that you would like the agent to access
• Structured Data - Specify any structured data sources that the agent would be able to execute queries against
• Response Instructions - Define the manner of responses that the LLM should generate

N.b. Any document retrievers or structured data sources you would like to use must be part of the project where this workflow is being built

The Tool Node has been improved to make finding relevant tools easier and more intuitive. The UI is split between Built-in Tools that have been build by Abacus and Created Tools that have been built by you or other members of your organization - you can also create new tools with the Create New Tool option.

Once added the node's configuration options can be accessed through the Node confiuration pane.

DeepAgent Workflow Builder A number of updates have also been made to how Abacus can autonomously build workflows using the DeepAgent. In order to access this functionality navigate to the Abacus front-end UI and under the LLM dropdown at the top of the page, select the AI Engineer:

From the options above the prompt window choose AI Workflow. DeepAgent will prompt you to provide specific details of the workflow that you would like to build and will then plan how to architect that as a workflow. After planning it will build out the workflow node-by-node, writing the relevant code snippets and testing as it goes:

After completion the workflow will be deployed and you'll be able to test it and make any additonal changes you need to all within the same conversation.

RouteLLM API Updates​

The RouteLLM API now supports PDF parsing and tool calling. Tools from OpenAI, Anthropic and Google are all supported, but requests should be made using the OpenAI API schema — Abacus will automatically translate requests into the correct schema for the provider.

For more details and examples please refer to the RouteLLM API documentation:

New Models​

Anthropic's Claude Opus 4.6 has been added to the platform. To find out how a range of new models perform don’t forget to check out our benchmark, LiveBench

Usage Throttling​

To help organizations manage overall use of the platform we have added the ability for admins to set daily usage limits. Navigate to the Profile >> Groups and Permissions >> Permissions menu:

Daily limits can then be set at either a global level or on a group-by-group basis. N.b. the limits apply to Super Assistant use only - i.e. RouteLLM, DeepAgent, Tasks. Autonomous agents or access of chatbots through our API's are not affected.

DeepAgent - Editing Function​

We have added a new Edit function for DeepAgent projects when building either apps or slides. One of the biggest challenges with these projects is that they often require edits and changes after the inital version has been built. Traditionally it is hard for the agent to pinpoint the exact context from the code to make only the relevant changes requested. The Edit feature solves this by providing an inline selection tool that allows you to pinpoint the exact changes you require. The precise context for that element is then passed to the agent making changes much more accurate and efficient.

To enable the Edit feature make sure you are in the preview window, then click on the Edit button at the top of the screen:

For powerpoint, Edit mode will enable you to edit inline elements of the slides - directly changing text and moving elements around.

With apps you will be able to select elements and then prompt the specific changes that you would like. The specific code for that element will be loaded into the chat as context to increase the efficiency and accuracy of the editing process

DeepAgent - Projects​

It is now possible to include DeepAgent conversations inside of projects, making it much easier to work collaboratively on DeepAgent tasks. To start a new DeepAgent conversation inside an existing project, using the Switch to DeepAgent mode button:

Alternatively you can add an existing DeepAgent conversation into a project:

Click on the ... next to a DeepAgent conversation and then use the menu to select Move to project

DeepAgent - App Updates​

A specific option has been added to the app management console to make scheduling database refreshes much easier. To access the App Management Console click on the Apps button in the top-left of the Super Assistant screen. Any apps that contain a database will give you the option the view or schedule

Clicking on the schedule button will open up a new conversation with DeepAgent to specify exactly what the refresh process should be. When complete the DeepAgent will create a Task that can be managed through the Tasks Management Console

You can use DeepAgent to create notifications for your deployed apps. These notifications can cover a range of things e.g. new users, completed forms, updated data. In order to access notifications you need to navigate to the app settings, which can be done in one of two ways:

• When previewing the app as part of the conversation - at the top of the page select Settings
• Through the App Management Console by clicking to View Versions or Database, or by clicking Manage under Custom Domain.

Once in Settings click on the Notifications tab

From the notifications screen you will be able to launch a DeepAgent conversation to build the specific notifications that you would like. Notifications are currently limited to email outputs, but the emails can contain any data that can be captured or analysed from the app. N.b. for more detailed analysis the app can be integrated with Google Analytics

In this example DeepAgent creates a simple email notification that tracks when the underlying LLM benchmark data changes with a summary of the new top models:

Another new option for DeepAgent apps is the ability to build multiple apps that leverage the same underlying database. This allows for much more efficient data management by preventing the need to host multiple copies of the same data. To do this navigate to the Settings page of the app that has the database you would like to re-use. On the Database tab click on the Create button in the top-right hand corner.

This will start a new DeepAgent conversation for designing the new app.

Finally, DeepAgent now supports building mobile apps. Simply add into your prompts that you'd like to build an app for mobile devices and it will optimize the build process, creating the frontend UI as well as all the backend api's that get deployed separately.

To deploy the app choose how the backend should be deployed (like you would any other DeepAgent app), and then use the Mobile App Packaging options to deploy the app as either Android or iOS:

OAuth Connectors​

We have extended support for generic OAuth connectors into the Super Assistant. Under Profile >> Connectors there is a new option for creating connectors into specific services using the Generic OAuth tool:

Clicking Add Config will open the configuration screen for authentication. Most services that support this type of connectivity will have OAuth or developer specific documentation that you will need to follow to generate the relevant IDs, urls and tokens.

A new option when building custom chatbots on structured data allows the integration more seamlessly between feature group and database data sources.

When configuring the chatbot select the FEATURE GROUP AND EXTERNAL DATABASES option in the Data Source dropdown. Then specify the feature groups and database connectors that the bot should have access to.

When asking questions of the bot it can reference both feature groups and database connectors in the same query.

Connector Updates​

We have added the ability for users of custom bots to be able to authenticate against multiple Sharepoint sites. A config connector for each site should be created and then added to the configuration screen of the custom bot.

Creating a config connector for the Sharepoint site is similar to creating a regular connector, except that you should select the config only option during creation:

Once your config connectors are created add them into the configuration screen for the custom chatbot. For each config connector the end users will be asked to authenticate the first time using the bot:

New Models​

OpenAI GPT 5.2 (including thinking and pro models) have been added to the platform. To find out how well the new models perform don’t forget to check out our benchmark, LiveBench

App Permissions​

Permission handling for DeepAgent generated apps has been updated to add more controls. At the organization level admins can set the default access permissions for all apps that are created. It can be managed in the Profile >> Groups and Permissions >> Apps menu.

• Private to Org - Apps are restricted to organizational users only
• Owner Only - Visible only to the creator/owner of the app
• Public - Open to the internet

Individual App Permissions The permissions for individual apps can be changed from the default by clicking on the cog symbol next to each app.

Within the settings for each app the owner or an admin can change the permissions between Owner Only and Private to Org. Only an admin can make apps Public if this is not the organization default.

When an app is Private to Org it is also possible to control access based on permission groups that have been created / are available in the platform. To do this use the Add Permissions drop-down in the app’s modify permissions screen:

Platform admins are also able to control who is able to share their apps, which limits an individual user's ability to assign permission groups to an app.

This is controlled through the Permissions menu (Profile >> Groups and Permissions >> Permissions). Adding groups into this selection means that users in these groups can only create apps visible to themselves

We have the first of a number of updates coming on our AI Workflow builder, with the overall aim to improve the ease of building workflows and to streamline the user experience.

The new workflow builder has a much cleaner UX to simplify the building process, which starts with the nodes themselves:

The code that defines each node is now hidden in the side-menu keeping the main workflow diagram cleaner and easier to understand. Clicking on any individual node will load the relevant code definitions on the right-hand pane.

Managing the node inputs and outputs has been improved. Clicking on the *Manage Inputs and Outputs *button on the side menu launches the input / output management pages.

1. NODE INPUTS Manage inputs that come from other nodes:

2. Easily identity from which nodes variables originate

3. *Required *nodes are clearly marked

4. Quickly convert to user input with the Convert to User Input button

5. Edit specific variables through the new configuration form

1. USER INPUTS Manage inputs that are from users:

2. Easily convert to node input, using the Convert to Node Input button

3. Required nodes clearly marked

4. Edit specific variables through the new configuration form

5. Manage overall input form structure and definition

6. OUTPUTS Manage node outputs:

7. Add / delete variables

8. Edit existing variables through the new Edit Outputs form

9. Visible in the UI variables are clearly marked to easily understand which are returned to users and which are passed through Node Testing

Testing individual nodes has been simplified by adding a Run button prominently on each node in the workflow diagram.

Enabling Run mode allows you to manually test individual nodes by providing the relevant inputs manually and evaluating the node outputs inside of the workflow editor.

Elements of the workflow configuration have been updated to provide much greater controls and options, especially when adding in connector nodes.

To access the configuration settings click on the cog symbol at the top of the page. In this example we have added the Slack tool to the workflow, and the configuration screen shows the actual scopes that have been enabled:

To update any of the specific connector scopes for these tools added to a workflow, click on the configure button. This will allow you to manage the specific access / capabilities for the connectors on a workflow by workflow basis.

More updates in the coming months…

Connector Updates​

A new Azure Storage Container connector has been added that can be authenticated using O-Auth. This can be configured to access a specific container or at account level

The generic O-Auth tool has been developed to make it possible for platform users to be able to create their own connectors to any service that supports O-Auth.

The configuration screen details the parameters that need to be provided to build the connector, as well as any scopes that you would like the connector to have access to:

Please note, the specific configuration parameters and scopes available will be determined by the individual service you are connecting to. Details of how to register and authenticate are often provided as part of providers developer pages.

The Outlook connector tools have been updated to also interact with Outlook calendar, enabling a far wider range of potential use cases

New Models​

OpenAI 5.1 and Gemini 3 have both been added to the platform. To find out how well the new models perform don’t forget to check out our benchmark,

Organization Level GitHub Policies​

A new setting has been enabled to manage organization access to the GitHub connector. To enable this navigate to the Profile menu by clicking on your profile on the top right-hand corner and selecting profile.

On the Profile page then use the menu on the left to select Manage >> Permissions