How It Works

The regression layer for production AI agents.

Capture agent state at failure time, propose a fix, and require replay verification before a pull request can open. Three steps with a defined contract: nothing merges until the replay passes.

Start Free Read the Docs

Step 1: Capture

Intercept the Exact Run That Failed

When your agent misbehaves in production, Relai's capture hook records everything needed to replay that run identically: tool call inputs, model responses, environment state, and the session context at the moment of failure.

  • 01
    Capture Hook

    A single SDK call wraps your agent's execution boundary and begins recording on any non-zero exit.

  • 02
    State Serialization

    Tool state, memory context, and intermediate outputs are serialized into a portable replay bundle.

  • 03
    Log Retention

    Configurable retention windows: 7 days on Developer, 30 on Team, 90 on Scale. Logs are encrypted at rest.

  • 04
    Replay ID

    Each captured failure gets a stable ID you can reference in PRs, Slack, or your issue tracker.

capture.log 
✓ relai.capture(agent_run)
  replay_id: rpl_0a1b2c3d
  captured_at: 2026-06-20T14:33:01Z
  tool_calls: 12 recorded
  model: gpt-4o-mini@2024-07
  exit_code: 1 (non-zero)
  context_tokens: 4,892
  state_size: 18.4 KB
→ Bundle stored. Use replay_id to reproduce.
Original Run FAIL 
✗ Step 4: summarize_findings()
  TypeError: expected list, got None
  at agent.run:line 88
  tool_state: {findings: null}
Fix Applied PASS 
✓ Step 4: summarize_findings()
  findings coerced to []
  tool_state: {findings: []}
  Replay passed in 1.2s

Step 2: Replay

Deterministic Replay on Every Fix Candidate

Every captured failure becomes a frozen environment. When a developer proposes a fix, Relai restores the exact tool state, model version, and context, then re-runs the agent against it. The replay is identical each time, eliminating the nondeterminism that makes AI bugs so hard to verify.

Step 3: Verify

Fixes That Have to Earn Their Merge

A fix proposal is only credible if it passes the failure it claims to solve. Before a PR can open, Relai runs the proposed change against the captured replay environment and gates the result. Green replay: PR opens. Red replay: the fix goes back to the developer, not to production.

diff: fix_null_findings.patch 
- findings = tool_result["findings"]
+ findings = tool_result.get("findings") or []
 
  replay_id: rpl_0a1b2c3d
  Replay Passed ✓
  PR gate: unlocked

CI Integration

Connect to Your Pipeline

Relai runs in GitHub Actions, GitLab CI, and any environment that can execute a shell command.

Agent Run Failure Detected Relai Capture Fix Proposed Replay Verified PR Opened
GitHub Actions 
- name: Install Relai SDK
  run: pip install relai-sdk

- name: Run agent with capture
  env:
    RELAI_API_KEY: ${{ secrets.RELAI_API_KEY }}
  run: relai run --capture ./agent.py

Everything Your Team Needs

Log Retention

Configurable 7, 30, or 90-day retention windows. Failure bundles are encrypted at rest and accessible via API.

Webhook Notifications

POST to any endpoint on failure capture and fix verification events. Integrates with Slack, PagerDuty, and custom receivers.

Multi-agent Support

Capture from parallel agent runs across distributed pipelines. Each run gets its own isolated replay bundle.

Diff History

Browse all fix attempts per failure ID. Compare how each proposed change performed on replay and track resolution time.

SDK Integration

Python and JavaScript SDKs. Install in under two minutes and start capturing with a single function call.

Status Dashboard

Per-agent health overview. See capture rate, replay pass rate, and mean time to fix across your entire agent fleet.