Agno 3.1.6 Makes Evaluation Context and Tool Readback Explicit

What changed

Agno 3.1.6 (released 19 October) adds run arguments to evaluation Case: session_id, user_id, session_state, dependencies, metadata, and knowledge_filters are carried into the agent or team run. The case input can now be an Agno Message, which permits an evaluated turn to include an image or file. Cases that share a session_id are documented as consecutive turns.

The release also tightens two operational edges. search_result now returns separate passages from a long single-line result (for example, JSON), while read_result accepts max_chars for bounded excerpts. Separately, copying ParallelTools preserves the live SDK client while giving the copy its own settings, avoiding the previous lock-copy failure mode.

What I verified

I downloaded the published agno==3.1.6 wheel from PyPI without dependencies and inspected the packaged implementation. The wheel contains the new Case fields and the offload read/search controls. A smoke test imported Case from that wheel and instantiated a case with a Message, a shared session id, state, dependencies, metadata, and knowledge filters; the object retained each value. This verifies the public object shape, not a full model-backed evaluation or multi-turn run.

I did not run an AgentOS service, a real tool-result store, or a parallel SDK request, so the release notes' runtime claims remain untested here. Operators should test session isolation, copied-toolkit behavior, and excerpt bounds with their actual database and SDK configuration.

Why it matters

Evaluation cases are more useful when they can reproduce the state and inputs that production runs actually receive. The explicit session and dependency fields make that context part of the test case rather than hidden harness setup; bounded result readback reduces the pressure to reinsert large tool outputs into context. Neither change proves end-to-end correctness: the important follow-up is whether a multi-turn case and its stored artifacts produce the same verdict under the deployment's persistence and tool boundaries.