Known limitations

During the public preview phase, API Agent has limitations for new chat sessions, chat interface settings, and login behavior.

New chat session

When you use the API Agent in Visual Studio Code, you can start a new chat session at any time. However, when you switch to a new session, you cannot return to or access the previous session.

Firefox login limitation

Logging in to the webMethods Hybrid Integration environment by using a private browsing window in Firefox is not supported.

Air-gapped installation

For this release, API Agent does not support an air-gapped installation.

Network accessibility issue

The API Agent runs on IBM’s remote infrastructure and needs to connect to the MCP server through the internet. If the MCP server is hosted on localhost, a private network, or any location that is not publicly accessible, the API Agent cannot reach it. As a result, local or private MCP servers do not work with the remote API Agent.

Authentication issue on Windows with Chrome and Edge

The API Agent currently only works with Firefox on Windows. When using Chrome or Edge, it does not redirect back to Visual Studio Code after login.

To authenticate and redirect successfully by using API Agent on Windows, complete the following steps:
  1. Set Firefox as the default browser on Windows.
  2. After authenticating, confirm and click the pop-up window that appears in Firefox.
  3. Manually return to the Visual Studio Code, and click the pop-up that requests permission to open the API Agent. This activates the Visual Studio Code plug-in.

Inconsistent responses in API Agent

The documentation search feature in API Agent is not automatic. The API Agent decides when to use the Documentation Searcher tool, and it does so if you explicitly request it.

The following scenarios explain why responses from the API Agent might vary:
  • The API Agent responds based on its internal knowledge instead of the most recent IBM Documentation.
  • Answers can be outdated, incomplete, or inconsistent.
  • Actively request a documentation search for accurate, version-specific information.
  • Newly added features might not be mentioned and deprecated features might still appear in responses.
  • The same question can produce different answers based on whether documentation was searched.

To get the accurate information, see the guidelines for using prompts.

Handling placeholder errors in single-operation plans

In cases where the API Agent's plan includes <place_holder> values, the first tool call can fail because placeholders are not allowed in single-operation plans. This violates the system prompt directive that the first tool call must be fully valid using only the user request and chat history.

To resolve this issue, take one of the following actions:

  • Modify the plan to provide valid arguments for all required inputs.
  • Rephrase the original request with more specific details so the agent can generate a valid plan without placeholders.

Mixed content and delays when finding similar APIs

When using the API discovery and rationalization tool to find APIs similar to a given API name, for example, get the APIs similar to risk-api, the API Agent response might include the following issues:
Table 1.
Issue Description
Latency The response can take more than 60 seconds to return.
Mixed content The output can include unrelated narrative text such as a comprehensive guide on API management combined with the actual tool output.
Duplicate results Tool output, such as the ArtifactSearcher table of matching APIs, might appear multiple times in the same response.
Misleading success message The API Agent might display a success message after plan creation, which can be interpreted as task completion. In reality, the plan is only built at this point, and you must run it to achieve the required outcome.
Impact
  • You can assume the requested action is complete when only the plan has been generated.
  • Mixed narrative content can confuse and obscure the actual tool output.
  • Duplicate results reduce clarity and can lead to incorrect interpretation.
  • Delays in response time can affect productivity.
Workaround
  • Verify run status: Use the Show steps view to confirm whether the API Agent has run the plan or only created it.
  • Proceed with execution: If the plan is generated but not executed, run the plan to obtain actual results.
  • Ignore narrative blocks: Focus on the structured tool output (tables) and disregard unrelated guide text.
  • Retry with specific queries: If latency is high or results are unclear, rephrase the request with more details, for example, include catalog name or API type.
  • Communicate clearly: A success message after planning does not mean the action is complete.