Two unattended days: gate completion on artifacts, not reports
Jul 20 Solar/Favur: six-session 2048 SOW run — false completion, oscillation, context exhaustion, review cost; validators cite docs/hashes; play-the-game signal beats green suite; task drift still open.
Source: What it takes to run an agent unattended for two days
Charles Solar / Favur (Jul 20) reports a two-day unattended run on Meta Muse Spark 1.1 against a 2048 SOW: 131 files with tests across six sessions, no human in the loop. The useful half is the failure taxonomy — better models raise output quality but do not remove long-loop failure modes. False completion: a step reports success without the work; validators gate advancement on artifacts (cite ≥2 context docs; reject commits without hash + non-empty file list; machine-check structured acceptance fields). Work and judgment are separated — the workflow agent sequences and enforces; tests are written before implementation. Non-convergence: oscillation between two wrong solutions looks like normal token spend; fix with per-step retry budgets that cannot re-arm, declared failure paths, re-entry caps, and run-wide circuit breakers. Context exhaustion is tier-summarized then hard-blocked near the ceiling. Review quality prices over-flagging with a false-positive breaker after three bad findings. Capability gaps may be abandoned without scoring penalty — otherwise agents conceal them with fabricated work. Checkpoints persist across the six sessions.
Honest non-goal: task drift has no semantic check; constraints are structural (fixed advancing tool, injected requirements, durable strategy from planning). Stronger completion than a green suite: the run launched and played the built game. Repo and full recording linked from the post.
Why it matters for code review
Always-on harnessed throughput compounds only when overnight “done” is artifact-backed and loops that oscillate halt instead of burning the night. Favur’s FOMO antidote is concrete — six sessions of verified slices beat a heroic Claude burst that still needs a human to notice false completion. Keep humans on drift and product promises; let the harness own hashes, budgets, and the refusal to advance on prose.