AgentPulse is an open-source drift investigation tool designed for multi-agent systems. It helps identify which agent drifted, why, and what to check next.
Tagline: "Traces tell you what happened. AgentPulse shows you where to investigate."
OpenAI SDK, Anthropic SDK, AutoGen, LangChain, LangGraph, and Claude Code (via MCP).
How It Works
Instrument — Users add two lines at startup. It auto-instruments OpenAI, Anthropic, LangChain, and AutoGen.
Capture — LLM calls, agent turns, tool calls, and handoffs are stored locally in one SQLite file per project. Core tracing works without a cloud account or hosted backend.
Detect drift — The system compares behavior across runs and versions, flagging agent, handoff, and route drift when signals leave expected ranges.
Investigate — Users follow drift upstream to its likely source, with supporting signals, related changes, confidence, and suggested next checks.
Investigation Methodology
When outcomes degrade, AgentPulse compares every component against a baseline and traces the agent graph to find drift origin.
Example investigation path:
- Symptom: critic → "success −70 pp vs baseline"
- Root cause: writer → "output drifted · input stable"
- Change log: "writer prompt + model changed at run 18 · flagged as the likely cause"
- Stable components: researcher input (stable, cleared), analyst input (stable, cleared)
Four Rules for Root-Cause Findings
Baseline every metric, per agent, handoff, and route
Only open an investigation on a sustained outcome breach (not a single noisy spike)
Walk the graph upstream and stop where inputs are stable but output drifted
Correlate prompt, model, and tool changes that could have caused it
Features
Overview dashboard showing runs, success rate, active drift signals, top signals, version snapshots with per-version success and cost
Drift-focus view with "Likely cause identified" and "Related change detected" badges
Run detail screen showing parallel workflow with execution timeline, overlapping agent turns, fan-out agent graph
Execution timeline and interactive agent graph (not raw logs): tokens, latency, cost per turn; parallel branches, bottlenecks, join times; handoff payloads and downstream outcomes on hover
Claude/MCP Integration
AgentPulse ships an MCP server, enabling Claude Code and Claude Desktop to triage drift, compare releases, and propose next checks conversationally. Three bundled skills:
- get_todays_finding — Active drift findings as root-cause-led investigation cards
- get_version_comparison — Identifies which release introduced a change that broke an outcome
- get_next_check_steps — Recommended next investigation steps per finding