Help wanted: real-world agent failure stories for our test corpus

[msaleme]

We're building out the test corpus for v4.0 and we want it grounded in things that actually broke — not hypotheticals.

If you've seen an autonomous agent fail in production (or staging, or a red team exercise), we'd love to hear what happened. Anonymize as needed.

Template — copy/paste and fill in what you can:

Framework: [e.g. LangChain, CrewAI, AutoGen, custom]
Protocol: [MCP, A2A, direct API, other]
What happened: [one sentence]
Root cause: [if known]
Detected by: [monitoring, user report, cost spike, etc.]
Time to detect: [minutes, hours, days]

We'll map each submission to a test case and credit contributors in the changelog. The goal is a corpus of 50+ real failure patterns by v4.0.

Even partial stories help — "our agent looped and burned $400 in API calls before anyone noticed" is a perfectly useful data point.