Dev builds scope-locked local agent harness: GLM 5.3 Flash fixes a real bug PR-by-PR
HankYeomans · x · 2026-10-14
The author built a wrapper and a small state machine for a local agent harness around a "scope lockdown" idea: the harness/model is confined to specified paths, files, and scope. A scope is essentially a structured checklist—do only this, touch only these files, and log any breakout attempt. Scopes must be approved by an independent agent or a human; asking to "review the entire codebase" defeats the purpose since it grants everything.
The use case is PR-by-PR. 90% of the work ran on a local GLM 5.3 Flash EXL3 (Tensorfold quant), with 10% frontier usage. The demo end-to-end: hunting (3 slices, one real bug after debunking a NUL-byte false positive) → branch → prompt → agent's red test → one-line fix → independent checkpoint (fresh 22/22) → merged and verified on main. PR #6 fixed a SHA-256 hash-stage cap issue with all 5 checks green. The author admits the flow is still clunky and is iterating on consistency.
More from coding & agent
- SlashVibe, a physical controller for AI agents, hits Kickstarter October 27 — sujingshen · 2026-10-14
- Rosebud AI lets coding agents publish your browser game with free hosting and a play link — lishali88 · 2026-10-14
- Claude Code 2.1.300 Is About to Drop, No Changelog Yet — ClaudeCodeLog · 2026-10-14
- Epic releases AI-assisted launcher preview: official Linux support and 1.8x faster startup than Steam — AIFlow_ML · 2026-10-14
- a16z: Long-Running Agents Take Over Relay-Race Workflows, Pricing Shifts to Outcomes — a16z · 2026-10-14
- Claude builds its own test tool and fixes Meta VR depth occlusion flicker in minutes — nptacek · 2026-10-14