How time-boxing worker agents works, explained with a four-quarter rule

I came across a post by Kaka Ruto, a developer who runs coding agents under a manager agent and had a familiar problem: some workers wandered the codebase and produced nothing. Their fix was a clock. Each worker gets one hour, split into four 15-minute quarters. The post reports "interesting results" but gives no numbers, so this explains the mechanism and does not judge that setup.

The rule

Quarter	Minutes	Rule
Q1	0–15	Work normally
Q2	15–30	No commit by the end: the worker is killed
Q3	30–45	Work normally
Q4	45–60	Not finished: stop and write a full report for the next worker

Ruto talks to one manager agent. The manager dispatches workers, reads what comes back, and uses it to brief the next ones.

Why it works

The rule turns an invisible failure into a visible one. A worker reading files in circles looks, from outside, the same as one making progress. Tying a deadline to something observable forces one of three outcomes within the hour: a commit, a stopped worker, or a written handoff.

The minute-30 check is a liveness test. The Bernstein project's design record says an idle agent is "indistinguishable from the agent being stuck, confused, or dead". A Claude Code issue describes a dead subagent still shown as running about 1.5 hours later.

The last quarter protects the explanation. A worker that runs out of time leaves code behind, but not an account of it. An issue in the ex_athena project puts it this way: "The work usually survives. The account of it does not." That project's fix was to stop each worker at 83% of its budget and ask for a report. The four-quarter rule reserves 25%.

The report feeds the next brief. Anthropic's post on its multi-agent research system says each subagent needs "an objective, an output format, guidance on the tools and sources to use, and clear task boundaries". A report from a worker that ran out of time is the raw material for a sharper one.

A handoff template

The post does not say what the report contains. This is a starting point I would suggest; it is untested against that setup.

Goal: what "done" means for this task
Done so far: what works now, and how I checked
Not done: what is left, in order
Tried and failed: approaches that did not work, and why
Read first: the files the next worker should open

"Tried and failed" saves the most time. Without it, the next worker is likely to repeat the same dead end.

Where it breaks

Four ways the rule fails, and a fix for each

Problem	Why it happens	A fix
A trivial commit passes the check	The rule tests that a commit exists, not what it contains	Have the manager check which files it touches
The minute-30 kill loses the report	Only workers that reach Q4 write anything	Ask for a short note before stopping a worker
Investigation tasks produce no commits	Tracking down a bug can take 30 minutes with no diff	Let a notes file count as the checkpoint
A fixed clock stops healthy workers	Time passing is not the same as being stuck	Use an idle timer that resets on progress

The last row has happened in practice. An ultraopen pull request says its earlier limit "killed every agent at 15 minutes — including legitimate agents that work for hours". The replacement was a 5-minute idle timer with a 4-hour ceiling. The second row matters most: a worker that could not make one change in 30 minutes has probably learned why the task is hard.

Takeaways

Tie deadlines to something the manager can observe, then check its content.
Reserve the end of every worker's budget for a report.
Treat the numbers as starting values. Nothing I found supports 15-minute quarters over any other split.

I have not run this rule myself yet. When I do, I will keep the last-quarter report as it is and change the minute-30 kill: an idle timer in its place, and a short note from the worker before it is stopped.

Sources

Kaka Ruto on X — "I started giving my worker agents an ultimatum" (2 October 2026)
Anthropic — How we built our multi-agent research system (13 June 2025)
Bernstein — ADR-005: Short-lived agents (22 March 2026)
ex_athena issue #237 — Reserve the end of a worker's budget for its handback (18 September 2026)
ultraopen pull request #17 — Make long agents survivable (14 September 2026)
Claude Code issue #54820 — Subagent liveness/heartbeat (29 April 2026)