Agent-run 'company' needed just 6 compactions in 1.5 months; paper shows benchmarks are compaction-friendly
maksym_andr · x · 2026-11-18
The author notes that Horizon Labs — the AI-run 'company' — operated for 1.5 months on a single Claude Code subscription with only 6 context compactions, plus a daily-updated COMPANYSTATE.md for major events.
The companion paper (arXiv:2610.31778) systematically studies compaction:
- Fixed OpenCode scaffold, 5 compaction strategies, 3 frontier models on PaperBench, ProgramBench, PostTrainBench
- For 1M-context models, compaction threshold has no measurable effect; GPT-5.3-Codex (smallest context) degrades at lower thresholds
- No method beats default self-summarization; more aggressive methods cost some models up to 20 points
- Risk: compaction silently drops task constraints — the first summary on ProgramBench loses 1/3 of anti-cheating rules on average, rising monotonically
Conclusion: most natural tasks are compaction-friendly; compaction + external memory effectively solves continual learning for agents.
More from coding & agent
- Raindrop ships MCP analytics: traces, issues, intent clustering for MCP servers — ycombinator · 2026-11-18
- New Codex plugin walks GPT-6.2 Astra through generating editable SVG and 3D assets — kieranklaassen · 2026-11-18
- 'Comprehension debt' is the new tech debt: never merge a diff you can't explain — therealdanvega · 2026-11-18
- FastMCP 4.2 ships Skills over MCP — drag a skill folder, get an MCP server instantly — HanchungLee · 2026-11-18
- Open-Source Local Agent Observer Locks Your Phone to Stop Doomscrolling — Roy3838 · 2026-11-18
- Amazon paper: weak LLM critics can sink search loops, ideal gain only 0.005 AUC — rohanpaul_ai · 2026-11-18