AgentPulse: drift investigation for multi-agent systems

Traces tell you what happened.<br>AgentPulse shows you where to investigate.

Open-source drift investigation for multi-agent systems: which agent drifted, why, and what to check next.

Get started free →<br>Chat with our engineers

Works with<br>OpenAI SDKAnthropic SDKAutoGenLangChainLangGraphClaude Code (MCP)

How it works

Two lines in, an investigation out

Instrument

Add two lines at startup. AgentPulse automatically instruments OpenAI, Anthropic, LangChain, and AutoGen.

from sdk import instrument<br>instrument(task_type='my-system')

Capture

LLM calls, agent turns, tool calls, and handoffs are stored locally in one SQLite file per project. Core tracing works without an AgentPulse cloud account or hosted backend.

Detect drift

AgentPulse compares behavior across runs and versions, then flags agent, handoff, and route drift when signals leave their expected range.

Investigate

Follow the drift upstream to its likely source, with supporting signals, related changes, confidence, and suggested next checks.

What you get

Built around the investigation, not the trace viewer

When an outcome degrades, AgentPulse does not hand you a dashboard of symptoms. It compares every component against your baseline and walks the agent graph to the point where the drift originated.

The investigation path

investigation walks upstream from the symptom<br>researcherinput stable · cleared

analystinput stable · cleared

Root causewriteroutput drifted · input stable

Symptomcriticsuccess −70 pp vs baseline

Change log: writer prompt + model changed at run 18 · flagged as the likely cause

Drift Investigation

Four rules turn a baseline breach into a root-cause finding:

Baseline every metric, per agent, handoff, and route

Open an investigation only on a sustained outcome breach, never 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 actually have caused it

Likely cause identified

Related change detected

See the run, not the logs

See every run as an execution timeline and interactive agent graph, not a wall of log lines.

Tokens, latency, and cost for every turn

Parallel branches, bottlenecks, and join time

Handoff payloads and downstream outcomes on hover

6.7s bottleneck branch

Claude integration

Let Claude run the investigation

AgentPulse ships an MCP server, so Claude Code and Claude Desktop can triage drift, compare releases, and propose next checks conversationally. Three ready-made skills are bundled.

> what drifted today?

1 drift finding across 3 projects

● writer · Agent behaviour · demo

Severity: Critical Confidence: High

Why: a change in writer propagated downstream

to critic, whose success worsened. The fix

belongs in writer, not critic.

Trigger: writer prompt changed near run 110

get_todays_findingActive drift findings as root-cause-led investigation cards

get_version_comparisonWhich release introduced the change that broke an outcome

get_next_check_stepsRecommended next investigation steps per finding

See it on your own agents

Book a 30-minute walkthrough of a real drift investigation, or ask how to wire AgentPulse into your stack.

Get started free<br>Chat with our engineers

Open experiment<br>AgentPulse is a reference implementation, and we want to learn how your team investigates agent failures. Tell us what works, what's missing, or how you do it differently.

Share feedback in Discussions →