AI AgentsOpt-in

Replay a real run before you change the prompt

Pick a step from a real execution, save it as a fixture, and replay it whenever the prompt, model or tools change. Replays run in test mode with every write tool removed and denied, so nothing reaches GitHub or Jira. Assertions on the output turn a suite of fixtures into a CI gate, from the app, a GitHub Action, or the MCP server.

Opt in: requires the agents.eval flag

• Fixtures saved from real executions, not hand-written transcripts
• Read-only replay: write tools stripped and deny-listed
• Output diff and claim checks between the original and the replay
• GitHub Action and MCP server run suites outside the app

1. Base agent typesSenior Developer, Code Reviewer, Tech Writer, Product Manager, and fully Custom agents.
2. Agent memoryPersistent memory tools let agents recall conventions, recurring issues, and decisions.
3. Multi-agent sessionsA coordinator delegates work across specialist agents and synthesizes the result.
4. MCP serversConnect remote MCP servers over streamable HTTP or SSE and their tools appear namespaced in every agent run.
5. Agent identity & ledgerOpt-inEach agent gets a stable identity, a signed short-lived token, and a ledger of everything it touched.