The Harness, Not the Model, Is What Makes an AI Agent Actually Work
alex_verem · x · 2026-09-10
A deep-dive on how agentic systems are assembled: an agent is not a model but a controlled workflow—trusted context + bounded tools + evaluation + human judgment + operational ownership. The model is the smallest, most swappable part; the recipe is instructions + scoped context + tools + a verifier + guardrails. Includes a worked example of a research agent that verifies who's asking, pulls only approved data, cites evidence, and stops when confidence is too low.
More from coding & agent
- Dev burns 2 days on basic email automation with Google Astra, blocked by CAPTCHAs and logins — AjdDavison · 2026-09-10
- Dev now clears a week of Codex and Claude output in about a day: 'honestly a little scary' — jacob_posel · 2026-09-10
- Open-Source Ref2V Workflow Auto-Transcribes Media and Formats Prompts for H3 Video Generation — bstr3k · 2026-09-10
- K3 got 77% speedup on mjwarp kernels; GPT-6 Astra added just 0.38% — YouJiacheng · 2026-09-10
- Same async program, four different outputs: paper maps the async/await design space across 7 runtimes — IanArawjo · 2026-09-10
- Dev says AI-written PRs keep getting smarter, and reviewing them is getting harder — smlpth · 2026-09-10