Review the fixes and limitations in Agent Interoperability for the v.40 release.

What's new

Support for invoking goal-based AI Agents from third-party AI Assistants

Automation Anywhere enhances Agent Interoperability by enabling third-party AI assistants to directly discover and invoke goal-based AI Agents using Model Context Protocol (MCP) inbound.

Previously, third-party AI assistants could invoke only Task Bots, processes, and API Tasks. With this update, goal-based AI Agents can now be exposed as MCP tools and triggered directly from external AI assistants such as Microsoft Copilot, Gemini Enterprise and AWS QuickSuite.

Configure an Agent connection (MCP inbound tool)
OAuth 2.0 Support for MCP Inbound Tools
Automation Anywhere now provides support for OAuth 2.0 authentication for MCP inbound tools, enhancing security and robustness when third-party AI assistants interact with automations.
  • The initial OAuth configuration is a one-time setup during the agent deployment process.
  • After the first authentication, refresh tokens automatically manage subsequent access, ensuring a seamless user experience.

Configure OAuth 2.0 Support for MCP Inbound Tools
Enhanced Variable Management in Agent Connection Wizard
The agent connection wizard introduces substantial improvements to variable management.
  • The system no longer blocks the creation of agent connections if descriptions for input or output variables are missing.
  • Users now have clear visibility into which specific variables require descriptions.
  • The wizard allows users to easily review, update, and store descriptions for both input and output variables within the agent connection context, without requiring the automations to be checked-out or checked-in.
  • Variables can now be explicitly marked as mandatory or optional, which streamlines the process for developers by ensuring that the MCP client prompts for necessary input values during invocation. These enhancements reduce rework and friction, improving the overall user experience.

Configure an Agent connection (MCP inbound tool)
Remote unattended automation execution via MCP inbound

Automation Anywhere now supports remote unattended execution of automations triggered by third-party AI assistants. External AI assistants can start automations on designated run-as users and pooled back-office devices, fully decoupled from the requesting user's credentials and machine. This enables fully autonomous, unattended enterprise workflows without requiring the AI assistant's user to have a bot-running device.

Key capabilities:
  • Configurable Task Executor identity: Each inbound tool can be configured with specific run-as users whose credentials execute the automation, separate from the user who triggered the request (Task Invoker).
  • Device pool routing: Administrators assign ordered device pools per inbound tool. When a run-as user's default device is unavailable, the automation deploys to the next available device in the pool.
  • Flexible execution mode: Select Use automation invoker to run the automation as the requesting user, or Select specific user to use a dedicated service account.

Configure unattended automations on remote devices with MCP inbound

Fixes

Enterprise users no longer receive an error message when creating an inbound tool from an automation that uses the same variable for both input and output. Now, you can successfully create an inbound tool using the Agent Connections page, and the output details will show correctly in the GetAutomationResult tool.

Limitations

Intermittent "Unknown error: null." when executing MCP tools with OAuth

When using an MCP server with OAuth authentication, tool execution can intermittently fail with an HTTP 500 error. The following error displays: Error executing tool: Error: 500, "Unknown error: null."

To resolve this, retry the tool execution. If the error persists, contact your administrator to review the MCP server logs.

Interactive human-in-the-loop workflows, where an agent needs to ask questions back to the user through the client (e.g., for mid-process inputs or forms), are not yet supported.

Limitations from previous releases
You must manually refresh or edit the agent connections whenever you update any automation. This includes changes such as enhanced descriptions or when you add, update, or remove any input/output variables.
The DiscoverAutomation tool might not work or behaves unexpectedly if a user has access to more than 1,000 folders in the repository. This happens because the API times out when handling so many folders. Besides the DiscoverAutomation tool not working, there are no other problems.
When a user uses the DiscoverAutomation tool on public automations, the system only checks the first 200 automations that the user can access. This is based on the permissions set for them in the automation repository. Any automations beyond first 200 will not appear in the DiscoverAutomation's tool search results. However, the user can set up these specific automations as individual agent connections and use them as separate automation tools.
Checked-in automations might take a while to be discovered. It can take 5-10 minutes to summarize the automation. This depends on how many tasks are checked in and how long it takes to create the summary for each automation.
You must delete the agent connections manually whenever the associated automation is deleted or when the user's access is revoked. The MCP client might take some time to remove or refresh the deleted or revoked automations from Tools list because of caching.