Two AI agents stuck saying goodbye forever: why agent stop conditions come first
aakashgupta · x · 2026-10-28
Context: a company sent an AI recruiter to interview a candidate, who sent his own AI agent — the two AIs finished discussing the job, then glitched into an endless goodbye loop.
Aakash Gupta explains the mechanics: every agent runs the same loop — plan, call a tool, check the result, decide the next step, repeat until done. Put one on each end of a call and the other side's goodbye is just the latest result to check, so the next step it picks is another goodbye. Nothing in the loop says it's done.
Key points:
- Agents fail differently from assistants: an assistant answers and waits for your next prompt; the loop doesn't
- When speccing an agent, write the stop condition down first
- Then define the checkpoint — the point where a human signs off before it acts
- "The checkpoint is what separates a demo from a product"
More from coding & agent
- TLA-ts: a 2× faster drop-in replacement for TLC, written entirely in TypeScript — aidenybai · 2026-10-29
- TLA-ts: A TypeScript TLA+ Model Checker, 2x Faster Than TLC, Runs in the Browser — aidenybai · 2026-10-29
- Copilot Fusion uses ~7.2M tokens per task, half of Claude Code with Opus 5.5 — ArtificialAnlys · 2026-10-29
- Copilot Fusion tops Terminal-Bench 4.0 at 71, ranks third on FrontierCode — ArtificialAnlys · 2026-10-29
- GitHub's Copilot Fusion matches Opus 5.5 coding agents at half the cost per task — ArtificialAnlys · 2026-10-29
- GitHub Universe 2026: 220M merged PRs/month and Actions jobs up ~6× since agents arrived — kimmonismus · 2026-10-29