Long-running agents suffer 'constraint amplification': a subtle form of context rot
generativist · x · 2026-10-08
@generativist shares a postmortem of a 'catastrophic run' from his agent projects, identifying a pernicious form of context rot in long-running agent sessions: constraint amplification.
- An initially narrow explicit constraint (per-experiment compute limits) gets gradually read as a broad goal;
- The agents then monitor PIDs, cgroups, disk, RAM and more, hard-failing on any excess;
- Constraints cascade and pile up, so the agent treats constraint-enforcement itself as the task.
A first-hand engineering lesson for anyone orchestrating long-horizon agents.
Related event: Long-Running Agent Sessions Suffer 'Context Rot' After Repeated Compaction(2 posts)→
More from coding & agent
- ezyang: types are good even for trivially testable things — no need to remember the tests — ezyang · 2026-10-08
- One engineer shipped web, desktop, iOS and Android apps in 6 months with AI — Yuchenj_UW · 2026-10-08
- Google Labs launches a 'GitHub for vibe coders' to store, share and remix AI-built apps — templecrash · 2026-10-08
- Keep coding agents sane: don't cram everything into one thread — msfeldstein · 2026-10-08
- Codex Cloud shipped with day-0 Tailscale support — and Tailscale didn't even know — pvncher · 2026-10-08
- Influzer ships MCP skill teaching agents to search, handshake, and paste configs — Fine_Airline_6832 · 2026-10-08