agentautopsy 2.0.0

Post-mortem debugger for AI agents

AgentAutopsy

Post-mortem debugger for AI agents — when your agent fails, this tells you exactly why.

---

Lifecycle

  ┌──────┐    ┌────────┐    ┌───────┐    ┌──────────┐    ┌─────┐    ┌─────────┐
  │ Fail │───▶│ Detect │───▶│ Trace │───▶│ Diagnose │───▶│ Fix │───▶│ Prevent │
  └──────┘    └────────┘    └───────┘    └──────────┘    └─────┘    └─────────┘
     │             │              │              │            │             │
  Agent        Pinpoint       Full SQLite     AI root      Auto-fix      Fix cache
  crashes      failure        event log       cause +      + replay      + drift
               step           locally         patch        verify        alerts

---

Commands

Command	What it does
agentautopsy ui	Open visual debugger
agentautopsy fix <id>	Auto-fix a failure
agentautopsy runs	List all runs
agentautopsy replay <id>	Print full event report for a run
agentautopsy mcp <cmd...>	Proxy and trace an MCP server over stdio
agentautopsy stats	Show fix-cache and token stats
agentautopsy serve	Start HTTP API (POST /analyze)

---

Features

Feature	What It Does	Use When
Zero-config watch	One import instruments LLM + HTTP calls	Adding tracing to existing agents
LLM interceptor	Records every OpenAI / Anthropic prompt and response	Debugging model behavior
HTTP interceptor	Captures failed outbound HTTP (http_error events)	Connection / API failures
SQLite trace store	Persists full decision traces in agentautopsy.db	Offline post-mortems
Failure detection	Finds the exact failing step in a run	Any agent crash
Root-cause analysis	AI diagnosis + concrete fix (Anthropic)	You need a fix, not just a stack trace
Fix cache	Remembers verified fixes for instant replay	Repeat failures
Auto-fix	Applies patch suggestions to your codebase	Turning diagnosis into code
Replay	Step through failed runs in CLI or UI	Understanding what went wrong
Web UI	Timeline, stats, debug chat	Visual debugging
MCP tracing	Proxy + schema drift / mismatch reports	MCP server or tool-call failures
Prompt diffing	Compares prompts vs. previous run	Silent behavior changes
Divergence detection	Flags runs that differ from past successes	Flaky or drifting agents
Multi-agent graph	Parent/child run chains (agentautopsy agents)	Crews and agent pipelines
GitHub Actions	Posts root cause + fix on PR test failures	CI/CD workflows
Slack alerts	Notifies channel on failure (AGENTAUTOPSY_SLACK_WEBHOOK)	Team visibility

---

Framework Support

Framework	Support Level
OpenAI	Native — auto-intercepted
Anthropic	Native — auto-intercepted
HTTP / httpx	Native — auto-intercepted
LangChain	Callback handler — get_callback_handler()
LangGraph	Callback handler — get_langgraph_handler()
CrewAI	Callback handler — get_crewai_handler()
MCP	Proxy CLI + watch_mcp() post-mortem tracing
Any OpenAI/Anthropic client	Works with zero config via watch()

---

Quick Start

1. Install

pip install agentautopsy

2. Instrument your agent

import agentautopsy

agentautopsy.watch()
# your existing agent code — nothing else changes

3. Debug failures

agentautopsy runs                  # find the run id
agentautopsy replay <run_id>       # see the full trace
agentautopsy ui                    # open the visual debugger

---

License

MIT — see LICENSE.