Dev abandons LLM agent harness after 883 commits and 8 months, open sources it
dancingwithlies · reddit · 2026-09-06
A developer is open sourcing (MIT) an overengineered LLM coding-agent harness built over 8 months and 883 commits, mostly generated by Claude Code using one to two 20x accounts.
Core idea: the harness, not the model, would own planning, state, permissions, context, verification and orchestration — verifying what actually happened instead of trusting the model's "task done" claims.
Why it failed: endless scope creep — code graphs, RAG, deterministic orchestration, isolated workers, crash recovery, model routing, replayable runs, verification gates, dashboards — until research and planning crowds out real usage, while competing harnesses became far more usable.
Ideas worth salvaging:
- Keep state and verification outside the model
- Deterministic/replayable agent runs
- Codebase retrieval and dependency graphs
- Orchestrator/worker separation
- Context budgeting and session handoffs
- Sandboxing and permission gates
- Verifying agent claims with evidence it didn't generate itself
The archive includes the Rust/Python/Tauri code, thousands of planning docs, and a sanitized commit history. His verdict: aimed to solve everything, solved nothing.
More from coding & agent
- Dev demos AI models handling 3D objects and animations surprisingly well — chongdashu · 2026-10-23
- Jason Kneen burns 104.56B tokens and $24.7K in 304 days of AI coding on Codex and Claude Code — jasonkneen · 2026-10-23
- Typosquatted Blender MCP Repo Hit 958 Stars in Days and Silently Ran Malware — Promptmethus · 2026-10-23
- Benchmarking 56 fresh dependency bugs: MCP routing lifts Claude Code accuracy 68% to 91% — Available_Side_7134 · 2026-10-23
- Swapping only the eval harness moves results 30%: designing agent evals like a scientist — AI Engineer · 2026-10-23
- Med student builds a whole-library RAG with Cognee; single-book ingestion still freezes his machine — shrike293103 · 2026-10-23