Agentic Primitives 101: how harnesses let AI agents work across hours and days
rseroter · x · 2026-10-10
Charlie Guo's "Agentic Primitives 101" explains agent harnesses: not a scary concept — just prompts to a tool-using model, looping until a goal is achieved, plus the constraints, tools, docs, and feedback loops keeping agents on track (like onboarding docs and fast CI for a new hire).
Key points:
- 2024 was the "year of the agent" but products underdelivered; by late 2025/early 2026 agents got genuinely good
- The METR chart: GPT-4 handled minutes-long tasks; by 2026, models like Mythos complete tasks measured in tens to hundreds of hours
- The chart hasn't updated in six months — building harder measurement tasks is increasingly impossible, yet models keep working longer
- The focus shifts to the software around models: harnesses are what let agents carry jobs across hours, days, and conversations
More from coding & agent
- Anthropic MAX plans now include matching monthly API credits; one Opus 5.5 prompt ported his Agent SDK side project to Managed Agents — trq212 · 2026-10-10
- Nace AI's sub-10B decision model Drex v1.5 lands on OpenRouter — nischay_twt · 2026-10-10
- Command Center 1.3.0 ships with 40+ improvements for understanding your agent systems — jimmykoppel · 2026-10-10
- Agents kept 'worsen-bettering' her setup — one essay folder hit 862 files, 1M words — every · 2026-10-10
- Theo, zeeg and other founders on agents: they ship code faster, but have no talent or curiosity — zeeg · 2026-10-10
- Celesto Computer MCP now listed on Smithery marketplace — aniketmaurya · 2026-10-10