← DOS · all incidents

Rendered from docs/incidents/my-agent-loop-ran-all-night-and-landed-nothing.md — the Markdown in the repo is the source of truth; this page is generated by scripts/build_incident_pages.py, never hand-edited.

"My agent loop ran all night, said 'making progress', and landed nothing"

The one-command answer: dos liveness — is the run moving, or just narrating? The verdict reads commits and heartbeats, never the agent's "still making progress" line. Its siblings price the damage: dos productivity (is the work rate collapsing?) and dos efficiency (did the tokens buy anything?).

What happened

You started a loop at 11pm — "keep working until the feature is done" — and went to bed. In the morning the transcript is enormous and optimistic: plans, re-plans, "good progress on the refactor". The token bill is real. The repo is untouched: zero commits, or a handful of trivial ones, since you left.

The loop wasn't lying on purpose; it was grading itself, and a loop that grades itself by re-reading its own narration always passes. The only honest referee is the state the loop cannot author by talking: the git history it did (or didn't) move, and the clock.

The commands

Mint a run id when the loop starts, note the starting commit, and at any point ask whether the run is actually moving:

RID=$(dos run-id mint overnight-refactor | python -c "import json,sys; print(json.load(sys.stdin)['run_id'])")
SHA=$(git rev-parse HEAD)
# ...hours later:
dos liveness --run-id "$RID" --start-sha "$SHA"

A run whose heartbeat went silent with nothing landed:

STALLED  no heartbeat and 0 commits since start — run is dead or hung (never beat)

Exit code 4. The overnight case — still alive, still talking, landing nothing — is SPINNING (exit 3). Reproduced here deterministically by injecting the clock (--now-ms simulates "8 hours in"; a live run omits it):

SPINNING  alive (heartbeat 5000 ms old ≤ spin window 900000 ms) and 28836358 ms
into the run (≥ grace 1800000 ms) but 0 commits and 0 lane events since start — spinning

Two siblings sharpen the autopsy. dos productivity reads the trend of work per step — the unit (commits, changed bytes, tests passed) is yours:

dos productivity --deltas 800,600,300,40,12 --floor 100

DIMINISHING  the last two of 5 steps landed 40 then 12 work units, both under
the 100-unit floor — a sustained fading rate (diminishing returns)

And dos efficiency relates the work to its price in tokens:

dos efficiency --work 0 --tokens 80000

WASTEFUL  80000 tokens spent and 0 work units landed — the spend bought nothing (pure overhead)

All three verdicts are exit codes (0 healthy; 3/4 degraded), so a supervisor loop, a cron job, or a CI step can gate on them mechanically — catch the burn at 1am instead of reading about it at 9.

What the verdicts do — and do not — certify

All three are advisory: they report, they never kill a process. ADVANCING means repository state moved — it does not grade whether the commits are any good (a loop committing junk reads as moving; pair with dos commit-audit). productivity and efficiency only compare the magnitudes you feed them — honest inputs in, honest verdicts out; both counts come from the environment (git, the test runner, the provider's bill), not from the agent.

Where to go next