Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

Knowledge

Agent loops that check themselves and keep going until done

How do you make an agent keep working until the result is there, check itself and improve, without you steering every step?

An employee views a work loop with a fault and a repaired part.
An agent works, checks and improves until the check passes. Illustration: Denkbaar.

An agent that asks “shall I continue?” after every step saves you no time. An agent that keeps working without checking makes mistakes you find later. You want a loop that works until it is done, and checks itself.

Plain technical English An agent loop repeats: work, check, improve. The loop only stops when a fixed check is green, or when the agent needs something that only a person can give. Each mistake becomes a rule, so that it does not come back.

Why it matters

You do not want to steer all the time. That is true for software, but also for a document, a deck or a marketing campaign. The loop must know by itself when the work is done, and when it needs you.

Work cards, a magnifying glass and a tool lie in a circle.
The loop repeats work, checking and improvement. Illustration: Denkbaar.

How it works

noyesyes

Goal and definition of done

Work

Check green?

Find the cause and fix it

Fresh reviewer

Done: link, image, test

Something only you can give?

One question to you

The agent works and checks until everything is green, and only stops for something only you can decide.

The idea comes from the factory: at Toyota, the line stops at a defect, the cause is fixed, and the process gets better (jidoka and kaizen). Google calls manual work that keeps coming back toil, and advises to automate it. And after a mistake, you write down what happened without blaming anyone (blameless postmortem).

A reviewer checks software on a laptop beside a phone and a printout.
A fresh reviewer tests the result. Illustration: Denkbaar.

What you do

  1. Say what “done” is. For us: a link you can open, a screenshot, and a test of at most six steps that ran green.
  2. Say when the agent must stop. Only for money, sending something outside, something irreversible, a new direction, or something only you can give, such as a key.
  3. No wait loops. We first had agents woken every hour by a fixed timer (a heartbeat); that did not get the work done, and we removed it. An agent reports a finished piece or a blocker, never “I am waiting”.
  4. On a mistake: the cause first. Only then the fix.
  5. A fresh reviewer who did not see the work tests the result.
  6. Learn: each correction becomes one line in a lessons file. The second time, it becomes a fixed rule or an automatic check. The third time manual work comes back, it becomes a script.

What we can do with you: we set up an agent loop for your task that works until it is done and checks itself. Book a conversation.

Sources

  1. obra/superpowers: verification-before-completion and other skills (MIT)
  2. Google SRE book: Postmortem Culture
  3. Google SRE book: Eliminating Toil

Topics: agentsmethodquality