This lesson is being recorded
Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.
Bruno Ashford @brunoashfordSubscribeModule 1 · Lesson 1.2 · 11 min
The Autonomy Ladder
Decide how much freedom to give an agent for each task: suggest, act with approval, or act alone.
Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.
1Rung 1: Suggest
The agent drafts or recommends, and you take the action. Use this when a mistake would reach a client, money or your reputation: replies to complaints, pricing, anything legal. The agent saves you the thinking time; you keep the final click.
2Rung 2: Act with approval
The agent does the work and asks before important changes. This is where most business tasks live: reports, proposals, content calendars. In Claude Code, the Manual permission mode works this way: it stops and asks before editing files or running commands, and you approve each one.
3Rung 3: Act alone
The agent works end to end and you review afterwards. Use it only for tasks where mistakes are cheap and easy to undo, like renaming files or tidying folders, and only after the task has worked well several times at rung 2.
4Use plan first for anything new
For a new or risky task, ask the agent to plan before it acts. Claude Code has a plan mode (you cycle permission modes with Shift+Tab) where it reads and proposes a plan without changing your files. You can also simply ask for it in plain English.
Before you change anything, read the files in this folder and give me a step-by-step plan. Wait for my approval before doing any of it.5Promote tasks slowly
Move a task up one rung only after three to five clean runs in a row. Move it back down the moment it makes a mistake that matters. This is exactly how you would manage a person, and it works just as well.
You would never give a new employee the company card and the client list on day one. You would start them on drafts, check their work, and give them more freedom as they earn it. Most owners forget this the moment the new employee is software, and they swing between two extremes: they either trust nothing or trust everything.
The Autonomy Ladder fixes that. It has three rungs, and every task you delegate sits on one of them. The rung is chosen by what a mistake would cost, not by how clever the AI seems.
Key points
- Three rungs: Suggest, Act with approval, Act alone.
- The cost of a mistake sets the rung, not how smart the AI looks.
- Most business work belongs on Act with approval.
- Plan first for anything new or risky.
- Promote slowly, demote fast.
Avoid
- Starting a brand new task on Act alone.
- Keeping everything on Suggest forever and wondering why nothing is saved.
- Approving every prompt without reading it, which quietly turns rung 2 into rung 3.
- Using the same rung for internal and client-facing work.
A 14-person practice put three tasks on the ladder. Client-facing tax deadline emails sat on Suggest: the agent drafts, a senior accountant sends. The monthly internal workload report went to Act with approval. Renaming and filing scanned receipts into client folders went to Act alone after two weeks of clean runs. Nothing client-facing ever went out unreviewed, and the team trusted the system because the rules were visible.