The AI delegation loop: hand a weekly job to an AI agent and stop fixing it in chat

You ask an AI tool for the weekly report. It is 90 percent right. You type the same three corrections as last week. Next week the chat is gone, and so are the corrections.

The AI delegation loop is a simple way out, from Actionable AI’s guide and John Bradshaw’s practitioner read of it: write the job down as a folder, let one general agent run it, and put every fix into the folder, not the chat. Here it is in five pictures.

1. One agent and a folder, not a bot per job

A specialist bot per job

• Its own prompt, tools and wiring
• The job changes, the wiring breaks
• You find out from a customer

Rebuilt every few weeks

One agent, one playbook per job

• A folder that describes the job
• Any good agent can read it
• Swap the model, keep the folder

The folder is what you keep

Two engineering teams point the same way. Vercel rebuilt its data agent from a pile of custom tools into one agent that reads well documented files: on 5 test queries it went from 274.8s to 77.4s on average, from 4 of 5 answers right to 5 of 5, with 37% fewer tokens and 42% fewer steps. Their own warning matters as much: it only worked because the files were already good documentation. Anthropic’s Agent Skills are the same idea in product form: folders of instructions, scripts and resources, instead of a custom agent for each use case.

2. The playbook: three layers in one folder

weekly-sales-report/

1. Process

   How is the job done?

   • Trigger and inputs
   • Steps, in your order
   • Decision rules
   • What done looks like
   • Odd cases
2. Toolbox

   What does it reuse?

   • The report template
   • A script that pulls the numbers
   • One good past example
   • Reference lists
3. Proof

   How do we know it is right?

   • Every figure traced to the sheet
   • Every link opens
   • Length under the limit
   • Names match the client list

• Process is how you actually do the job, written for a smart new hire: when it starts, what goes in, the steps, the rules for the choices, what finished looks like, and the odd cases.
• Toolbox is what should not be reinvented each run: templates, scripts, reference data, one example of a good result. Without it you get a slightly different result every week.
• Proof is a short list of checks that must pass, with evidence, before anything reaches you.

3. A check that cannot fail is not a check

The agent reads its own draft

• “Looks good”
• “All figures are correct”
• “Links checked”

You still inspect everything

Checks with evidence outside the draft

• Each number matched to a cell in the sheet
• Each link opened, status shown
• Word count against the limit
• Each claim matched to a quote

You read the evidence, not the whole draft

Research backs this up: models are poor at catching their own reasoning mistakes without outside feedback (Huang and colleagues, ICLR 2024). Anthropic’s advice for agents is the same: prefer checks that code can run, and judge the real outcome (does the booking exist in the database?), not the agent’s sentence saying it booked. Without proof, you delegated the draft and kept the inspection.

4. The loop: fix the folder, not the chat

The playbookprocess, toolbox, proof

1 Run2 Check3 Find the layer4 Fix the folder

1. 1Run The agent runs the job from the folder, from scratch.
2. 2Check The proof list passes or fails, with evidence.
3. 3Find the layer A person asks: missing step, stale tool or missing check?
4. 4Fix the folder The smallest lasting fix, in that layer. Then run again.

When the output is wrong, name the layer before you fix anything:

• It skipped a step you always do. That is the process: add the step or the rule.
• The format drifted, or a number came from an old file. That is the toolbox: update the template or the script.
• A wrong figure reached you. That is the proof: add the check that would have failed.

5. How far along are you?

1. 1Chat fixesCorrections live in a thread that closes.Resets weekly
2. 2Giant promptEverything in one block that keeps growing.Brittle
3. 3Written processSteps, rules and odd cases on paper.Repeatable
4. 4Process and toolboxTemplates, scripts and examples saved.Consistent
5. Quack!zz5Closed loopChecks and notes feed the next run.Compounds

Start this week

1. InputPick a weekly jobOne you would spot a bad version of instantly.
2. AI stepGet interviewedLet the agent ask questions and write PROCESS.md.
3. RuleRun once, freezeSave what came out right into the toolbox.
4. OutputAdd checksFive to ten that can fail, with evidence.
5. PersonRead the evidenceThen run it again. The second run is where it clicks.

A folder for the weekly sales report could look like this:

weekly-sales-report/
  PROCESS.md   steps and rules
  toolbox/     template, script
  proof/       CHECKS.md
  NOTES.md     every fix, dated

Any agent that can read files can use it: Claude Code reads a folder like this directly, and Agent Skills package the same idea.

When not to do this

• The job happens once. A playbook pays off on the second run, not the first.
• Your files are a mess. Vercel’s result depended on good documentation; a folder of stale notes gives you faster wrong answers.
• Nobody owns it. Someone has to read the evidence and fix the right layer. That decision stays with a person.

Where this shows up

Hoàng Huy Đức describes the work as the automation layer under marketing and sales, from lead scoring to CRM pipelines (LinkedIn, September 2026). See AI agents for how a first playbook is scoped.