The Coding Agent Core Loop Is Simple ReAct — the Hard Part Is Harness Engineering
blaizedsouza · x · 2026-11-03
A reposted thread argues that a coding agent's core loop is just the classic ReAct pattern — Reason → Act → Observe → Repeat: explore the codebase, edit files, run tests, read errors, retry. The loop itself isn't the hard part.
The engineering challenge is everything around it:
- Managing context
- Handling permissions
- Recovering from failures
- Knowing when the task is complete
That's where building a reliable coding agent gets complicated — and where harness engineering begins. The accompanying diagram lays out a full Coding Agent Loop architecture: a TUI with a steering queue for user goals and interventions, a headless harness containing the LLM↔Tools loop plus a context window (system prompt, goal, tool calls), and an outer cycle of Tasks Plan / Explore / Apply / Execute phases, with TODO Write continuously updating the plan until the goal is achieved.
More from coding & agent
- Google & HKU's VeriCUA cuts false task-completion rate to 5.4% by checking saved files, not agent claims — rohanpaul_ai · 2026-11-03
- GitHub ships wave of new Copilot features post-Universe, Fusion GA with 30% off — mariorod1 · 2026-11-03
- Laurence Moroney publishes 'Fine-Tuning AI' with O'Reilly, betting on fine-tuning careers — lmoroney · 2026-11-03
- Direct Weight Editing via Codex+Astra Lifts Audio Classifier Accuracy From 84.3% to 88.75% Without Retraining — dansitu · 2026-11-03
- baoyu-design adds Motion: run Claude Design locally with real-time animation editing — dotey · 2026-11-03
- Honest Thumbnails: Chrome extension swaps YouTube clickbait covers for real video frames — jdjohnson · 2026-11-03