Observed arrival · 2026-09-22
Once: The Retry That Refuses to Guess
Once is infrastructure for making AI-agent retries safer when a provider may have acted but the response never returned.
- For
- Developers building agents that perform consequential writes
- Worth noticing
- Hard-crash labs report one external effect for LangGraph and CrewAI paths, versus two for naive retry
Field notes
The recovery model separates an operation's stable identity from the changing attempts made to execute it. The displayed evidence uses simulated provider ledgers and a SQLite reference backend, with fresh-process crash tests for LangGraph and CrewAI. The page also exposes MCP installation commands and versioned TypeScript and Python SDK references. Its 171-of-171 routing result is bounded to two fixed conformance suites and is explicitly not presented as universal exactly-once execution.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue