Reading and Writing Carry Different Risk

An agent that reads records fails in one way: it produces a wrong report. An agent that writes to records fails in another: it changes something other people rely on. The first failure costs time. The second failure alters the record, and someone has to find the change and reverse it.

Every agent design starts with one question: does this agent read or write? The answer decides how much control the agent needs around it.

Starting Below Auto-Execute

Auto-execute is the appealing tier, since it removes the reader from the loop. It also removes the feedback that shows where the agent errs. An agent that starts by acting has no record of correct proposals behind it, and its first mistake lands directly on someone's work.

Starting lower costs some speed in the first weeks. The speed returns later, because each action that moves up the ladder arrives with evidence behind it.

Report-Only

A report-only agent reads and describes. It has no path to change anything. Its worst outcome is a bad report, which costs a reader a minute. The Overseer works this way, and the Program Overseer skill is a template of it.

New agents start here. A report-only version shows whether the agent reads the right things and reaches sensible conclusions. That evidence gathers before the agent gets any ability to act.

Propose-Only

A propose-only agent works out what it would do, writes the proposal where I can read it, and stops. Each proposal names the action, the reason, and the evidence the agent used. I approve or reject each one, and nothing executes without approval.

The cost is my time. The benefit is visibility: I see the agent's reasoning, and I learn where it errs. Proposals have to stay short, since a design that needs long reading stops getting read, and an unread proposal queue is an agent with no supervision.

Auto-Execute

An auto-execute agent acts without waiting for me. This tier suits actions that can be reversed, that affect little, and that the agent has proposed correctly many times. An action reaches this tier after a history of approved proposals.

The tier belongs to each action. One agent can propose some actions and execute others, and each action sits at the tier its own record supports.

An Independent Reviewer

For agents that execute, a second agent reviews each action before it runs. The reviewer runs in a fresh context and receives the proposal and the evidence. It forms its own view and approves or rejects.

The reviewer is independent for a reason. An agent that reviews its own work tends to approve it, because it starts from the same assumptions that produced the proposal. A separate reviewer starts from the evidence.

An Iteration Cap

When the reviewer rejects a proposal, the first agent can revise and resubmit. This loop needs a fixed cap. Without a cap, two agents can pass a disagreement back and forth indefinitely and burn time and money while producing nothing.

When the loop reaches the cap, the item goes to me. The cap turns an unresolved disagreement into a human decision, which is where an unresolved disagreement belongs.

A Hypothetical Ladder

Consider an agent that sorts incoming requests into categories. At the first tier it reports how it would sort each request, and I compare its choices with my own. After a period of close agreement, it moves to propose-only: it assigns each request a category in a proposal, and I approve each one. If approvals arrive unchanged over a long run, the category assignment can move to auto-execute with an independent reviewer checking each action. Reassigning a request to a different owner stays at propose-only, because other people see that change.

The example shows that each step depends on evidence from the step before it. The agent receives each capability only after it demonstrates the same skill at a lower tier.

A Record of Every Action

Any agent that executes needs a log. Each executed action gets recorded with the proposal, the evidence, and the reviewer's verdict. When something goes wrong, the log shows what the agent did and why it believed the action was right. The same log supports the periodic review of each action's tier.

Choosing a Tier

Three questions place an action on the ladder. Can the action be undone? Who sees the result? How often has the agent proposed this action correctly? An action that cannot be undone or that other people see stays at propose-only or below. An action that is easy to undo, visible only to me, and repeatedly proposed correctly can move up.

Moves happen one tier at a time, and any surprise moves the action back down. The ladder runs in both directions.

The final post collects what Claude and I learned across these programs.