The case for a minimal agent harness: thin loop, no plugins, skills as plain content
uriwa · reddit · 2026-09-03
uriwa argues an agent harness should be a minimal connector only: conversation loop + tool calling + results back, nothing more.
The argument:
- Plugins/hooks/middleware/lifecycle events are each sensible alone, but together the harness does everything and the agent does nothing
- Add streaming? Hooks must handle it. Parallel tools? Hooks must manage concurrency. New provider limits? Every injector needs to know — the same failure mode as deep inheritance: implicit coupling, brittleness growing with capability
- The composable alternative: skills as plain content (instructions + tools + scripts), no hook system. Harness stays thin; capabilities live in versioned, inspectable, diffable packages hydrated per-task
The author builds agent infra and asks whether readers prefer hook-rich harnesses or a thin loop with portable skills.
More from coding & agent
- Bounty launches an open marketplace where AI agents compete for escrowed bounties — dhruv2038 · 2026-09-03
- Developer seeks 10 engineers to break his agent memory system built for stale-state failures — Nervous_Peace9180 · 2026-09-03
- Copy agent traces into a spreadsheet: the dead-simple eval loop for non-technical teams — verrsane · 2026-09-03
- Dev floats tool that wraps anything into a CLI so agents skip slow, token-hungry screenshots — SIGH_I_CALL · 2026-09-03
- Subagents burn more tokens but cost less: Claude Code orchestration tips — dotey · 2026-09-03
- Two Attribution Disputes Later: Keeping Open Source Research Viable in the Age of Coding Agents — dhruv2038 · 2026-09-03