The Weft Manifesto

One number: five

No more than five finished results or open questions wait for your decision at once — ten when a team shares the board. When the queue is full, an agent that asks to start something new is told to finish, shorten or pick up returned work instead. Finishing is never refused: finished work is never thrown away.

The limit is small on purpose. In Kingman’s queueing formula, waiting time grows with ρ/(1−ρ): a reviewer who is busy 95% of the time keeps work waiting 19 times longer than one who is busy half the time, with the same variability. Like Google SRE’s goal of keeping toil below 50% of an engineer’s time, five is a number you can break — and see that you broke it.^7,8

Seven principles

1. 01Work is no longer scarce. Your decision is.

   Put the hard limit on the queue in front of the person who decides, not on the Doing column. An idle agent is a healthy agent; it does not get tired, and three tasks or fifty are the same to it. The person reviewing its work is the one resource that is used up one item at a time.

2. 02Every “done” says what it rests on.

   An agent’s report is an application, not evidence. In single-control τ²-bench domains, false success — the agent says done, the environment says otherwise — makes up 45–48% of failures, and no LLM judge reading the agent’s trace scored above an AUROC of 0.65, where a coin flip scores 0.5. So the strongest “done” reads the outside world: a page that answers, a file whose hash matches, an object that exists. Work that needed nothing from you may close on the agent’s word — and is then shown as exactly that, never as checked.⁴

3. 03Whoever did the work does not sign it off.

   Verified and accepted are two different statuses. A server that re-read the outside world, a deterministic check or a person can mark work verified; the agent that did it cannot. Accepting is a second, separate signature. For anything irreversible — money, publishing, a message to a client, anything legal — Weft asks the agent to stop and wait for you, and once that stop is set, no agent can lift it or walk the task into Done.

4. 04Every task has one human owner.

   You cannot hold an agent to account; you can only assign it. So a task has exactly one owner, who is a person, and a delegate, who is named: not “the AI”, but Claude Code, Cursor, ChatGPT or Codex. Keep the delegates few. Anthropic’s own advice for agent teams is to start with three to five, because “three focused teammates often outperform five scattered ones.”⁵

5. 05At the limit, route — and never fake a decision.

   When the queue is full, the system says what to do instead of throwing an error. When an agent goes quiet mid-task, its claim lapses and the task returns to the board for someone else. And when you have not decided yet, the work waits for you: silence is not acceptance, so nothing is ever marked accepted because time passed.

6. 06Nothing carries over silently.

   The week is the unit of planning: people were 62.9% more likely than on an ordinary day to commit to a goal at the start of a week, and 23.6% more likely at the start of a month. A week ends with a decision on every unfinished task — continue with a concrete next step, split it, send it back or drop it — rather than with a quiet move to next week.⁶

7. 07The decision outlives the session.

   No AI client hands its memory to anyone else, and it is right not to. So what survives a session is what was written down outside it: not a file on you, but the history of the work — how a task ended, what was decided and why. Memory has to forget, too: a note that is out of date stops being given to agents.

Run your board through this

A manifesto without a measuring stick is a poster. These seven questions take about five minutes to ask of any task manager. Where Weft has not closed a property for every user yet, it gets a dash, not a tick.

✓ works today · — partly, or not yet switched on · ✗ not there

What we count, and what we do not

Velocity, story points, burndown, utilisation, the number of tasks created and the number of AI calls are not in this product. They measure a resource that stopped being scarce, and they reward noise.

A feeling of speed is not a measurement either. In METR’s study of experienced open-source developers using early-2025 tools, they believed AI had made them 20% faster; measured, they were 19% slower.⁹

What is worth counting is four things: how long the oldest item has waited for you; the time from “the agent is ready” to “you decided”; the share of work accepted the first time, separately for each executor; and the minutes of your attention per accepted result. The last one is the only honest test of whether AI is paying off. Weft records what each of them needs — who did each piece of work, when it was ready and what you decided — and shows how much is waiting on you.

Where this breaks

Some work has no executable check. A letter to an investor or a positioning statement cannot be verified by a server, and there the method honestly falls back to an ordinary board with one reviewer and one such task at a time. A framework that promises more than it can check breaks exactly where someone relied on it.

There is no peer-reviewed study yet of the right limit when agents do the work. Five is our default, not a finding. We started at three and raised it after the first weeks of real use: a third of the refusals came with only three or four items waiting, and they were cleared in one sitting. Review may also get cheaper: GDPval itself notes that review time falls as professionals get used to the work. If checking ever becomes cheap and reliable, the limit should grow — which is why every threshold here is a default we expect to revise, not a belief.³

Sources

1. 2026 Software Engineering Benchmarks Report — LinearB, 2026
2. The Acceleration Whiplash — Faros AI, 12 April 2026
3. GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks — OpenAI, arXiv:2510.04374, 5 October 2025
4. From Confident Closing to Silent Failure — arXiv:2606.09863, 1 June 2026
5. Orchestrate teams of Claude Code sessions — Choose an appropriate team size — Anthropic, retrieved 22 October 2026
6. The Fresh Start Effect: Temporal Landmarks Motivate Aspirational Behavior — Dai, Milkman, Riis — Management Science 60(10), 2014
7. The single server queue in heavy traffic — J. F. C. Kingman — Proceedings of the Cambridge Philosophical Society, 1961
8. Site Reliability Engineering, Chapter 5: Eliminating Toil — Google, 2016
9. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR, 10 July 2025