# AI Delegation Loop Source: John Bradshaw — https://bradshaw.cloud URL: https://bradshaw.cloud/strategy/ai-delegation-loop/ Type: insight Date: 2026-09-18 Tags: AI, AI agents, Automation > Stop building a specialist agent for every job. Capture process, toolbox and proof in a playbook a general agent can run — and put every correction back into the folder. ## The playbook model The playbook framing below is adapted from Actionable AI's AI Delegation Loop guide at theactionableai.com — it is not mine. What follows is my read for practitioners who keep rebuilding the same agent wiring. ### Specialist agents die. Job knowledge does not. One agent per job carries its own prompt, tools and wiring. When the job changes, the wiring breaks and you find out in production. What a general agent lacks is rarely intelligence — it is your judgment written down as a playbook it can run. **Business take:** Treat the playbook as the durable asset. The model is replaceable; the process, toolbox and proof for a weekly job are not. ### Process, toolbox and proof are different jobs The process layer is judgment in the order you actually use it: triggers, inputs, steps, decision rules, definition of done and edge cases. The toolbox holds reusable scripts, templates, reference data and examples. The proof layer is a checklist that can fail with evidence outside the draft. **Business take:** Teams that only write a prompt are still doing quality control themselves. Teams that skip the toolbox rebuild the same template every run — and get a slightly different result each time. ### A check that cannot fail is not a check Self-review — the model reading its own draft and declaring it fine — is worth nothing. A real check produces evidence outside the text: a figure traced to a source file, a live link, a count against a limit, a spelling matched to a reference list, a claim matched to a quote. **Business take:** Without proof you have not delegated the job. You have delegated the draft and kept the inspection. That is how “90 percent done” becomes your permanent role. ### Fix the playbook, not the chat When output is wrong, diagnose which layer failed: a missing step (process), a rebuilt or stale artifact (toolbox), or a missing check (proof). Make the smallest durable fix in that layer, note it, and re-run from scratch. The chat closes; the folder compounds. **Business take:** Correcting in the chat trains you to type the same fix every week. Correcting in the folder is how the second run needs less of you than the first. ## From chat fixes to a closed loop Maturity is whether the next run needs less of you than the last one. ### Delegation maturity How durable the job knowledge becomes - Chat corrections — Fixes live in a thread that closes — Resets weekly - Giant prompt — Everything in one growing instruction block — Brittle - Written process — Purpose, steps, decisions and edge cases exist — Repeatable - Process + toolbox — Templates, scripts and examples are frozen — Consistent - Closed loop — Proof checks and notes feed the next run — Compounds _Start with the job you do weekly and would recognise a bad version of instantly. Do not start with the most complicated thing you do._ ## Delegation decisions ### Should this job get a playbook? **Only if it repeats and a bad version is obvious.** A good first playbook is something you do at least weekly, would spot as wrong in seconds, and could explain to a new hire in under twenty minutes. Skip one-off research and anything you cannot yet recognise as finished. ### What belongs in the toolbox? **Anything a new hire would reuse on day one.** Scripts for mechanical steps, templates with placeholders, reference spellings and IDs, finished examples you were happy with, and rules files for tone or format. Leave out one-off outputs, secrets and drafts you have not approved. ### Is this check real? **Only if evidence lives outside the draft.** Trace a figure to a source file, open a live link, match a claim to a quote, count against a limit, or check a spelling against a reference list. The model declaring its own work "professional" is not a check. ### Where do I put the fix? **In the layer that failed — then re-run.** A missing step is process. A rebuilt or stale artifact is toolbox. A mistake no check would have caught is proof. Write the smallest durable change into the folder, note it, and start the job again from scratch. Do not leave the lesson in the chat. ## Frequently asked questions ### What is the AI delegation loop? A way of handing a repeated job to a general AI agent by capturing it as a playbook — process, toolbox and proof — and putting every correction back into that playbook. The loop is run, fail, diagnose the layer, fix the folder, re-run. The chat closes; the folder compounds. The framing comes from Actionable AI's AI Delegation Loop guide; this page is my practitioner read of it, not a reprint. ### Why not build one specialist agent per job? Each specialist carries its own prompt, tools and wiring. When the job changes, the wiring breaks and the operator finds out late. Actionable AI's guide points at the same failure Vercel and Anthropic engineers have written about: chaining specialists looked clever, then a general agent plus a folder of the team's own files outperformed the wiring. What was missing was job knowledge, not another agent. ### What are the three layers of a playbook? Process is judgment written down: purpose, triggers, inputs, steps, decision rules, definition of done and edge cases. Toolbox is the reusable scripts, templates, references and examples the process points at. Proof is a set of checks that must pass with evidence outside the draft before anything reaches you. ### How do I start in thirty minutes? Pick the weekly job you recognise a bad version of instantly. Have the agent interview you into a process file, run the job once and freeze whatever came out right into the toolbox, then add five to ten checks that can fail. The second run is where the pattern clicks — stop only after you have seen it. ## Diagrams Standalone SVG diagrams, licensed CC BY 4.0 with attribution to John Bradshaw — link back to https://bradshaw.cloud/strategy/ai-delegation-loop/ when embedding. - [Specialist agents die. Job knowledge does not. (SVG)](https://bradshaw.cloud/strategy/ai-delegation-loop/diagrams/playbook.svg) - [Process, toolbox and proof are different jobs (SVG)](https://bradshaw.cloud/strategy/ai-delegation-loop/diagrams/layers.svg) - [A check that cannot fail is not a check (SVG)](https://bradshaw.cloud/strategy/ai-delegation-loop/diagrams/proof.svg) - [Fix the playbook, not the chat (SVG)](https://bradshaw.cloud/strategy/ai-delegation-loop/diagrams/loop.svg) ## Related - [Agentic AI patterns](https://bradshaw.cloud/strategy/agentic-ai/) — choose the architecture before you widen autonomy. - [Agent authority](https://bradshaw.cloud/strategy/agent-authority/) — sponsor, mandate and evidence before an agent can act. - [Model Context Protocol](https://bradshaw.cloud/strategy/mcp/) — the tool connection layer a playbook will call. - [AI Delegation Loop guide](https://theactionableai.com/guides/ai-delegation-loop/read) — Actionable AI's original playbook framework this page reads from. - [AI governance in practice](https://bradshaw.cloud/writing/ai-governance-in-practice/) — who owns the guardrails when software takes action.