-
Notifications
You must be signed in to change notification settings - Fork 4
Governance As Code
A policy document has never stopped an agent. A checked gate has.
Governance without velocity is bureaucracy. Velocity without governance is a liability.
Most organisations have the first one and call it the second one's fault. The way out is not less governance. It is governance that executes rather than governance that is published.
| Written governance | Governance as code | |
|---|---|---|
| Where it lives | A document, a policy site, a slide | The deployment pipeline, the gateway, the agent's own configuration |
| How it is enforced | Someone remembers | The thing does not run |
| How you audit it | Ask people | Read the log |
| What happens under deadline pressure | It is waived, verbally | It is waived with a name and an expiry date, or it is not waived |
The last row is the one that matters. Every governance system gets overridden. A good one records who overrode it and when the override expires. A bad one has no record because there was nothing to override.
What the agent must never do, written as specific outcomes rather than as principles.
These are never events: not "be careful with pricing" but never promise a price that is not in the database. Not "respect professional judgement" but never override the named professional's decision.
A never event is testable. That is the entire difference between this and a policy document. If you cannot write a test that fails when the agent does it, it is a value, not a never event, and it belongs somewhere else.
What the agent cannot do, because the system will not let it.
Input and output controls: sensitive-data detection before anything reaches the model, output scanning for the failure patterns you have actually seen, and injection defence anywhere an outside party can put text into the system.
Layer 2 is the only layer that works when nobody is watching. Layers 1 and 3 depend on someone reading something.
Who is accountable, and what they are accountable for.
A named owner per agent. A review that happens on a date rather than on a feeling. An escalation path that was written before the incident.
Three layers, and the common failure is having only the first. A policy layer with no technical enforcement and no named owner is a document that describes a system nobody built.
| Pillar | What it means in practice | The rule that makes it real |
|---|---|---|
| 1 · Visibility | Every agent has a registered ID, a named business owner, a risk tier, and a link to its audit log | No registry entry, no API key |
| 2 · Security | Sensitive-data detection on every input before the model sees it; output scanning; injection defence on anything externally facing | Enforced at the gateway, not requested in a guideline |
| 3 · Quality | Every agent has a golden dataset built by a domain expert, not by the developer. The pipeline blocks a deployment when accuracy falls below the baseline | The block is automatic. A quality gate a human can wave through is a quality suggestion |
| 4 · Cost | Cost per transaction reported to the sponsor on a cadence. Automatic suspension when cost or error rate breaches a threshold | The threshold is set before the agent runs, not after the invoice |
Pillar 4 is the one that gets added after the first surprise invoice, and it is the cheapest of the four to build up front.
| Circuit breaker | Kill switch | |
|---|---|---|
| What it does | Degrades the agent automatically: back to read-only, back to draft mode, scope blocked, spend capped | Stops the agent |
| Who triggers it | A threshold. Error rate, cost spike, drift signal | A person |
| When it fires | Whenever the threshold is breached, including at 3am with nobody awake | When someone decides |
| Why you need it | Because the failure you did not anticipate will happen outside working hours | Because some failures should not be degraded, they should end |
A programme with only a kill switch has a control that requires a person to be paying attention, which is exactly the assumption agentic systems break. Build the circuit breaker first.
Design the kill switch to be fast, obvious and testable. A global hard stop that revokes tool permissions and halts the queue, plus softer switches for session pause and scoped blocks. Test it on a schedule. An untested kill switch is a belief about a kill switch.
An immutable log of prompts, retrieval, reasoning and outcomes.
It is the thing that makes every other control auditable, and it is almost always the thing that was not built, because it produces no visible feature and costs storage.
The test for whether you have one: an agent did something wrong six weeks ago. Can you reconstruct what it saw, what it retrieved, what it decided and why, without asking anyone? If the answer involves a person's memory, there is no system of record.
The repo publishes a pre-flight checklist with ten gates. Source material behind this page describes seven.
The published ten are the artifact of record and they are what this page endorses. The discrepancy is recorded here rather than reconciled silently, because a governance page that quietly changes the number of gates is demonstrating the exact failure it is warning about.
One rule about all of them: gates are checked, not claimed. A README asserting something exists is not evidence that it does.
The alternative to enforcement is not freedom, it is a shadow estate.
The CoE provides a standard platform: infrastructure, logging, guardrails, the registry. Teams that build on it get faster approval and support. Teams that go off-road own their own risk, explicitly and in writing.
This is what makes governance survivable. A governance function that only says no gets routed around. One that offers a faster path for the compliant route gets adopted, and the exceptions become visible rather than hidden.
From the practice canon. Developed through client delivery and refined in collaboration; the frameworks stated here are the author's own.
The thinking
Operating model
Frameworks
Governance
Playbooks
Value and people
Reference
In the repository
The courses
Reviewed 2026-08.