Skip to main content

Python Agent v2.2.1

September 29, 2026

LLM Observability​

  • New Added support for AWS Bedrock Runtime instrumentation

    • Collects the Converse, ConverseStream, InvokeModel, and InvokeModelWithResponseStream calls of bedrock-runtime as LLM API steps

  • New Added the ability to stop an active trajectory

    • You can stop a running trajectory on the Active LLM screen

      • The in-flight LLM call is cut off and the transaction ends. LLM calls raised afterward in the same trajectory are blocked as well

    • The stop point is recorded in the transaction profile as a WhaTap: LLM trajectory stopped by control message step, so you can tell it apart from an application error

  • New Added the llm_blocked_users String option

    • Default: ""

    • Blocks LLM calls from the given user IDs. Separate them with commas; case is ignored.

      • Example: llm_blocked_users=u-1024,abuse@example.com

    • A blocked call returns as a connection error of the SDK, such as APIConnectionError, and a WhaTap: LLM call blocked for user … message step is recorded in the transaction profile

  • New Added the llm_httpc_step_enabled Boolean option

    • Default: true
    • Collects HTTP calls made inside an LLM workflow, such as MCP and REST tool calls, together with the LLM steps
    • Set it to false to collect only LLM API calls

  • New Added a metric for newly started trajectories

    • The active trajectory metric now includes the number of trajectories started during the last five seconds (new_count)

  • Fixed Fixed LLM API steps and API status code statistics not being collected on openai 3.x and anthropic 1.x

  • Fixed Fixed LLM transactions being lost in short-lived processes

    • Transaction end information and LLM logs were not sent from processes that run once and exit, such as CLI runs and child processes

Common​

  • New Added instrumentation for the httpx2 library

  • Changed Streaming HTTP responses now measure elapsed time up to the point the response ends

    • For streaming responses such as SSE, the step elapsed time runs until the response completes instead of stopping when the headers arrive

  • Fixed Fixed the native agent starting once per worker on Linux when whatap_code_start was set to true in a multi-worker environment such as uvicorn --workers

Note

Agent download: whatap-python v2.2.1

Last updated on