Integrations

If you already score your agent with another tool, Trace can tell you whether the difference between two versions is real. poisson-trace-import reads the result files that tool wrote, one import per version, and poisson-trace fix-check compares them.

poisson-trace-import deepeval old.json --store ./store --run old \
    --metric "Tests Passed" --cut recorded --case name
poisson-trace-import deepeval new.json --store ./store --run new \
    --metric "Tests Passed" --cut recorded --case name
poisson-trace fix-check old new --store ./store --event missed --inputs all --max-mde 0.05

You name the metric, the cut and what identifies a case. Nothing has a default. The importer runs no metric and calls no model.

Evaluation tools

Tool	Reads
DeepEval	DeepEval’s saved test runs, one metric per import
Ragas	The score table saved from Ragas, one score column per import
promptfoo	The results file promptfoo eval -o writes (JSON, JSON Lines or CSV)
Inspect	Inspect’s eval logs
OpenAI Evals	The records an eval run writes
Braintrust	Experiment results exported from Braintrust
Weave	Evaluation results exported from Weave
LangSmith	LangSmith runs
OpenTelemetry	OTLP trace exports
Lemma	Lemma’s trace payload, captured on your machine
JUnit	The JUnit XML your test runner or CI writes
Any table	A CSV or JSON Lines file your tool or script saved

Every reader is set up with your team during onboarding, against the files your stack already writes.

Agent frameworks

Poisson’s runtime runs an agent built on one of these frameworks and captures its trace while it runs, so the stored runs carry the agent’s steps, model calls and tool calls.

Framework	instrument	Spans stored per episode
OpenAI Agents SDK	["openai_agents"]	agent 5, model call 2, tool call 1
LangGraph	["langchain"]	agent 6, model call 2, tool call 1
CrewAI	["crewai", "openai"]	agent 2, model call 2, tool call 1. Both instrumentors are needed

Compute

Prove runs on your machine by default. For a large sweep it runs on a cluster in your own AWS, GCP or Azure account, set up with your team during onboarding, or on Modal (poisson-prove sweep --backend modal).