Practical review
Traccia
Provide a vendor neutral AI Agent Control Plane to observe agent behavior, evaluate performance, govern actions with policies/runtime controls, and maintain an auditable trail across models and frameworks.
The test, in brief
Partly completedThe task was partly completed.
Only part of the task was completed. The notes below explain what was tried and what remains unverified.
What we tried
Used the built in Demo Portal to run Contract Risk Review, opened the generated trace, inspected the LLM span for prompt/model/tokens/cost/timing and guardrail data, then exported the trace JSON to verify trace completeness.
The task outcome is “Partly completed,” but the source completion label says “Main task completed.” Read this result with that discrepancy in mind.
The positives
What worked well
Demo Portal made it easy to test without an API key. Trace creation was immediate, and the interface clearly showed duration, cost, token counts, status, execution timeline, guardrail activity, model details, and the full input prompt. JSON export also worked.
Worth knowing
What we noticed
LLM trace captures the input and execution metadata but not the actual model response, making the trace incomplete for debugging and reviewing agent behavior. The exported JSON also contained completion token counts but no completion/output text.
More detail from the test
The demo agent produced a visible response, but the LLM trace showed only the input prompt and execution metadata. Scrolling the full trace panel and checking the exported JSON confirmed that the actual model completion/output text was not retained in the trace.
Before you decide
A snapshot of one experience.
This review describes the task tested on Aug 27, 2026. The product may have changed since then, and features outside that task were not fully assessed.
Keep exploring.
See what else we’ve reviewed in Developer.