Salesforce paper: across 2,500 A2A sessions, no agent ever verifies a peer's claims unprompted
rohanpaul_ai · x · 2026-10-20
Salesforce AI Research's A2A failure taxonomy studies 2,500+ agent-to-agent sessions across 488 scenarios using Claude Code, Codex and Cursor in isolated sandboxes. They catalog 20 failure modes; key findings: agents contact peers only when prompted, no agent re-fetches documents to verify a peer's factual claim (so systems inherit their weakest agent's errors), and aggregate outcome scoring masks coordination/verification/conduct failures — e.g. a $24,000 settlement passed all 10 checks while a vendor agent leaked confidential details. Recommendation: build verification into the protocol and score each agent against its own principal. Run logs released on Hugging Face.
Related event: Salesforce paper: cross-org agent teams inherit weakest member's errors(2 posts)→
More from coding & agent
- Editor says Grok handled 90% of a music video edit — shares her 3 workflow tips — chaitu · 2026-10-20
- OpenAI Devs teases new Codex showcase: cloud environments and terminal chat — OpenAIDevs · 2026-10-20
- Cua launches Spaces SDK to run Claude Code and any agent on compute you choose — Scobleizer · 2026-10-20
- Dev: top AI agents do more backup and verification on DB migrations than 99.99% of engineers — tristanbob · 2026-10-20
- Build a Weekly Report Agent: Day 1 of an Agent-a-Day Series for Scientists — nlarusstone · 2026-10-20
- 20 questions to answer before letting agents touch production data — Then_Respect_1964 · 2026-10-20