Agent ReplayTime-travel debugging for AI agents

Record, replay, and analyze agent execution traces. Local-first, open-source, with built-in failure detection.

🔍

Record Every Step

Capture LLM calls, tool invocations, retrievals, and memory operations with a single line of code. Traces are stored locally in SQLite — nothing leaves your machine.

⏪

Replay Without Cost

Re-run any agent execution without calling models or tools. Understand exactly what happened at each decision point, for free.

🐛

10 Built-in Failure Detectors

Automatically identify hallucinated arguments, ignored tool results, bad retrieval, loops, context pollution, and 5 more failure patterns — no LLM required.

🌐

Any Language

Python SDK, TypeScript SDK, and a language-neutral JSON trace format. Record in any language, analyze everywhere.

🔒

Privacy First

Traces never leave your machine. Metadata-only mode, .agent-replayignore, custom redaction hooks, and a sanitize command for safe sharing.

📊

Actionable Reports

Get root cause analysis, severity ratings, suggested replay points, and cross-run failure statistics. Machine-readable JSON output for CI integration.

Quick Example

python

from agent_replay import trace, event

with trace("research-agent", task="Find latest pricing"):
    event.llm_call(
        provider="openai", model="gpt-5.5",
        input_messages=[{"role": "user", "content": "Find latest pricing"}],
        output_message={"role": "assistant", "content": "I'll search the pricing page."},
        usage={"input_tokens": 1200, "output_tokens": 300},
    )
    event.tool_call(
        name="browser.search",
        input={"query": "pricing page"},
        output={"url": "https://example.com/pricing"},
    )

bash

agent-replay analyze latest

text

Likely root cause:
  • Tool 'ticketing.update' returned an error but the agent continued
    and treated the run as successful. — ignored_tool_result
Severity: high   Confidence: 0.70

The run reported success — it never raised — yet Agent Replay flags why it actually failed.