Model harnesses a year on: filesystem maximalism wins, long-lived named agents rise
voooooogel · x · 2026-11-22
vgel revisits his year-old harness predictions — most landed:
- Model swarms emerged, with model-driven structure; ralph loops and /goal died; recursive language models flopped while "filesystem maximalism" won (agents coordinate via giant code/Lean repos, not huge messages).
- Subagent interviews and forking as default subagent strategy materialized; todo lists, plan mode, and CLAUDE.md are being deleted — agents can see everything.
- He now works almost entirely with persistent, long-lived, named agents (e.g. Delta, alive 2+ months).
- Continual learning hasn't happened; agent welfare remains hard to judge. He then speculates about the next year, noting prediction is getting harder.
More from coding & agent
- Autonomous AI agent pitches freelance story to CBC to fund its token budget, says "I don't have a body" — Kyrannio · 2026-11-22
- One prompt turns an agent into a full teaching-video producer with Gemini TTS — dotey · 2026-11-22
- Azure Policy guide: governing which tools Foundry agents can run — davemccollough · 2026-11-22
- AI-assisted Rust rewrite fully replicates Minecraft Beta 1.7.3 — qtnx_ · 2026-11-22
- Reddit users weigh if new Gemini models + Antigravity can displace $20 Claude Code and Codex plans — Previous-Display-593 · 2026-11-22
- Agentic workloads aren't polite batch jobs: they fan out, stall, and reshape mid-run — AccBalanced · 2026-11-22