Notes · Design ·
Authority before action: governing agents like a sales organization
Most agent governance reads like an IT policy. The better model is one every sales leader already runs: territories, qualification, pipeline review and approval levels.
This is a design, not a deployment. Parts of it run in my own setup today: scoped workers, a software verifier, receipts, read-only guards. The organizational model around them is the proposal.
The problem
An agent can do far more than it should be allowed to do. The research on agent governance this year keeps landing on the same question: not "can it?" but "is it authorized to, here, now?" The Authority Resolution Framework (August 2026) treats that as a reasoning problem of its own: who may act, on what, and why.
Sales organizations solved a version of this decades ago. A rep can work any account in their territory, but a discount past a threshold goes to a deal desk, and a contract goes to legal. Nobody calls that bureaucracy; it's how a team moves fast without one person's mistake becoming everyone's.
The mapping
| In a sales organization | For a team of agents |
|---|---|
| Territory | Scope: the paths, tools and data an agent may touch for this task, and nothing else |
| Qualification criteria | An acceptance contract, written and frozen before work starts |
| Pipeline review | Agent review on a cadence, from receipts, not from the agent's own summary |
| Deal desk and approval levels | Graduated authority: act alone, act and notify, or wait for a yes |
| Forecast discipline | Calibrated confidence: a typed judgment with a probability, not a paragraph |
| Win/loss review | Every failure becomes a check or a rule, so it can't repeat silently |
Three levels of authority
| Level | Examples | Why |
|---|---|---|
| Act alone | Read, search, draft, run tests, write to its own branch | Reversible and contained |
| Act, then notify | Open a pull request, change a non-production setting | Reversible, but someone should know |
| Wait for a yes | Merge, deploy, spend money, contact a customer, delete anything | Hard to undo, or outward-facing |
Evidence a stranger could audit
Approval is only as good as the record behind it. The Decision Evidence Maturity Model (April 2026) grades whether the evidence an agent leaves is enough for an outside party to reconstruct a decision. My rule is the same, in plainer words: a decision without a receipt didn't happen. The receipt names the commit, the checks that ran and what they returned.
What my commercial experience contributes to AI teams
My sales and partner work makes me attentive to ownership, review cadence and the points where a decision needs another person's judgment. I bring that operating perspective to collaboration with engineers: define who owns the decision, what evidence they need and when the work should stop or escalate.
The organizational model described here is a proposal. Some underlying controls run in my own setup; the broader operating design still needs evaluation in a team environment.
Give agents territories, not permissions lists; contracts, not prompts; and a deal desk for anything you can't undo.