AI agent governance · v1.0
AI agent governance, from someone who runs it.
AI agent governance is who can make an agent act, what it may do without asking, and how you prove what it did. If the third part is missing, the first two are paperwork.
My companies run on AI agents: seven model harnesses pulling from one shared task queue on a two-hour schedule, around the clock. Anything that leaves the building, an email, a post, a payment, waits at an approval gate for me. These are the twelve rules that system enforces, with the numbers it produced.
Why most AI governance fails
Most governance guides are policy documents: principles, risk tiers, a committee. None of that runs at 3 a.m. when an agent decides to retry. What I learned the hard way:
- An agent posted to my LinkedIn without asking. I deleted it myself.
- A publishing API saved each post, then returned an error. The agent retried. 17 duplicate articles went live.
- A bot sent a prospect two emails by mistake, then suggested fixing it by turning the approval gate off. I turned it back on.
- Thirty-nine approved drafts sat unsent for three weeks because nothing owned the step after approval. A gate without a decider stops delivery too.
Every one of those was a missing rule in the runner, not a missing paragraph in a policy. The full ledger is here.
The three layers
One owner per task, a locked queue, a kill switch that stops everything.
Anything that cannot be undone waits for a person. Consent is checked before contact.
An outcome log nobody edits, a receipt per task, cost per attempt, counts you publish.
The twelve governance rules
One agent is responsible from claim to verdict. A handoff does not complete a task.
Every settlement is appended to a log nobody edits. The board is derived from it live, never a set of pushed cards.
Queue and board writes are locked merges. One agent never overwrites another's status.
I use 30 minutes per task and one pause file that stops every runner.
Send, publish, spend, provision and migrate wait for a person. Approval is not delivery. Testing never removes a hold.
Do-not-contact and consent checks run before the action, not after.
A held run is the gate working. It is also zero deliveries. Count both.
Tokens and dollars per run, so you can compute cost per verified task instead of cost per claim.
When a loop breaks, a watcher queues its own repair task. Green exit codes prove nothing.
Expired auth or a usage limit releases the task back to the queue and rotates to a healthy harness.
Any change that weakens a check lands with the case that justified it.
Proof rate, hold rate, cost per verified task, revenue you can trace. Anecdotes are not a scorecard.
These rules are enforced by a control plane, not a policy document. Here are its nine parts.
Governance is not observability
| Orchestration | Routes work between agents and tools. | Who does the next step? |
| Observability | Shows traces, logs and latency after the fact. | What happened inside the run? |
| Evals | Tests an agent against cases before you ship it. | Is it good enough to deploy? |
| Governance | Decides what may happen and proves what did. | Was it allowed, and did it really happen? |
You can have perfect traces of an agent doing the wrong thing. Governance is the only one of the four that asks whether the action was allowed and whether it really happened outside the system.
The scorecard, with my own numbers
| Metric | Formula | Mine, Oct 1 to Nov 6 |
|---|---|---|
| Proof rate | done at P3+ ÷ done claimed | 1 of 585 with enforced read-back |
| Hold rate | held ÷ settled | 223 of 1,436 |
| Cost logged | attempts with cost ÷ attempts | 8%, wired up Nov 2 |
| Verified revenue | dollars traced to P4 receipts | $0 |
Bad numbers, published on purpose. The next version reports cost per verified task in dollars.
"Proof rate" comes from the definition of done for AI agents: a task is done only when something other than the agent reads the result back from the system where it lives. Governance without that definition counts claims, not outcomes.
Govern your agents this week
- Write down every action your agents can take that cannot be undone: send, publish, spend, provision, delete. Put a person in front of each one.
- Give every task one owner and one outcome, written before the work starts.
- Add a time cap per task and one switch that stops all of them.
- Make every run append one line to a log nobody edits: what was claimed, what proof came back, what it cost.
- Run the free checker on that log and get your proof rate. Publish it, even if it is bad. Mine was 1 of 585.
The checker is one Python file with no dependencies: agent_dod.py. Instructions are on the definition of done page.
Questions
What does AI agent governance entail?
Three things: control over who can make an agent act, limits on what it may do without a person approving, and proof of what it actually did. Policies on paper are not governance until the system running the agents enforces them.
Who controls an AI agent?
Whoever owns the task and the runner. In a governed setup every task has one named owner, actions that cannot be undone wait for a person, and a single kill switch stops every agent. The agent itself never approves its own outward actions or grades its own work.
What is an AI agent control plane?
The layer that decides what agents may do and records what they did: the queue, the approval gates, the kill switch, the outcome log and the checks. It is the place where governance rules are enforced rather than written down.
How is AI agent governance different from observability?
Observability shows what happened inside a run. Governance decides what is allowed to happen and proves the outcome in the real world, such as the email in the sent log or the charge in the payment processor.
Do you need an AI agent governance platform?
Not to start. You need the rules enforced where your agents run: a gate on irreversible actions, one owner per task, a time cap, a kill switch and an outcome log. A platform helps at scale, but it cannot replace a definition of done.
Rule IDs are stable, so link to them directly, for example /ai-agent-governance#gov-5. Text under CC BY 4.0. If you want these rules running inside your company, get in touch.
Get the next scorecard.
Real numbers from agents running a real company, including what broke. One email when there's something worth reading.