// the find
lidangzzz/goal-driven
A multi-agent system that keeps running for ~100 hours and solve a very complicated coding or math problem that can be verified
goal-driven is a prompt template, not a library — it's two README files describing a pattern where a 'master agent' repeatedly checks if a 'subagent' (run via Claude Code, Codex, etc.) has met a verifiable goal, restarting it if it stalls or claims false completion. It's aimed at people who want to point an agent at a huge, well-specified, checkable problem (port a compiler, prove a theorem) and let it grind for 100+ hours unattended.
The core supervisory loop is a genuinely useful pattern for long-horizon agent work: separating liveness checks from goal-verification, and restarting a stalled worker instead of trusting its self-reported 'done', sidesteps a real failure mode of autonomous agents. It's grounded in actual output, not just theory — linked example repos (a C++ TypeScript compiler, a Rust SQLite reimplementation) show the approach produces something runnable. The emphasis on criteria that can be mechanically checked (diff compiler output against tsc, for instance) is the right instinct, since it's the one thing that keeps the loop from rubber-stamping itself.
There's no code in the repo at all — just two markdown files; 'usage' means copy-pasting a prompt block into your own agent tool, so this is a technique writeup, not software. No cost or safety controls: a loop explicitly told to run 100-300 hours and 'not stop until the user stops it manually' with auto-restarting subagents is a direct path to runaway token/API spend, and nothing here addresses budgets, logging, or a kill switch. The 'master agent checks criteria' step is itself just another LLM judgment call with no guidance on wiring real automated verification (test suites, diffing, etc.) — which is the actually hard part of making this safe to trust. The pseudocode only shows the happy path; there's no discussion of what happens when a subagent loops on a misread of the criteria or burns hours making no progress.