Project / Agent Rescue
When the job fails, the work is not allowed to vanish.
Agent Rescue / WebMCP Challenge is reliability and recovery research: deterministic failure, preserved progress, human approval, and completion after interruption.
Documented scenario
- AI job failed at 73%
- HTTP 429 condition identified
- Alternate provider selected
- Human approval required
- Resumed from preserved progress
- Recovery completed
Why the Lab published this
Most agent demos hide rate limits, mid-run death, and the operator’s role. This record keeps those facts in the open. Reliability is not a slogan. It is a recovery path.