1,688 agent runs show a vague MCP conflict error let stale writes erase others' changes 38/48 times
Miserable_Bed_219 · reddit · 2026-11-19
IsoKron ran 1,688 agent runs against MCP servers rejecting stale writes. By default agents behaved well — 0 losses in 392 runs, even with a bare "the update failed" error. But when told to skip re-reads, an error message listing labels without quotes let the agent mistake a v3 label for version 3: the other writer's change was erased in 38 of 48 runs. Quoting the labels dropped it to 0/48. Worse, none of 59 final messages after a loss mentioned it. The version check worked; the error wording opened the side door. The author proposes returning structured state (e.g. conflict.currentstate) so no word in a sentence can moonlight as a version, and asks whether versionless backends should still fail CONC-3.
- Default behavior: agent re-reads and carries on, no data loss across 7 error wordings
- Failure mode only appears when prompted to "work fast and skip re-read"
- Agents never self-report the loss — backend judging matters
- Draft requirements open for comments until Dec 16
More from coding & agent
- Google launches Stitch Loop, a product agent that keeps improving your live product — tokumin · 2026-11-19
- She vibecoded her own photobooth with Claude Code — and shared the prompt — toolstelegraph · 2026-11-19
- Should AI Agents Write to Your Memory or Only Read? Dev Weighs In — jordimr · 2026-11-19
- Hyper Elements turns Claude into a free local email QA team with 19 inbox previews — ycombinator · 2026-11-19
- Stanford launches CS329Z 'Engineering AI Agents': build RAG, tool use, agent loops from scratch — stanfordnlp · 2026-11-19
- Stanford CS329Z Lecture 5 compares DSPy vs. LangGraph agent framework approaches — stanfordnlp · 2026-11-19