Governance

Agent Autonomy Controls

Give every agent exactly as much freedom as it has earned — and not a step more. Autonomy controls decide, on every single action, whether an agent moves on its own or pauses for a human to say yes. One dial, set by the people who own the risk.

The one idea Autonomy governs a single question: when an agent is about to change something in the real world, does it go ahead, or does it wait for approval? Everything below is how you tune that answer per agent, per team, and across the whole organization.

Reads are free. Writes are the moment that matters.

An agent looking things up — searching a knowledge base, reading a ticket, checking a calendar — carries no risk of changing anything, so it never needs permission. The moment that matters is the write: booking the trip, filing the ticket, sending the message, deleting the record. Autonomy controls sit exactly on that line.

Kind of action	What it looks like	Governed by autonomy?
Read	Search, look up, summarize, retrieve	No — always runs freely
Write	Create, update, send, book	Yes
Destructive	Delete, cancel, overwrite, irreversible changes	Yes — held to the highest bar

Three levels of trust

Every agent is set to one of three levels. Think of them as how long a leash you hand the agent — from "check with me on everything" to "I trust you, go."

Moving left to right hands the agent a longer leash — from pausing on every change, to only the dangerous ones, to none at all.

Supervised

The safe default. Every change waits for a person to approve it. Perfect for a brand-new agent, a sensitive workflow, or any time you want a human in the loop on all of it.

Trusted

The agent handles everyday work on its own, but still stops before anything destructive — a delete, a cancellation, an irreversible change. The best of both: speed on the routine, a checkpoint on the risky.

Autonomous

The agent runs end to end without pausing for autonomy's sake. Reserved for proven, well-scoped agents where the rest of your guardrails already carry the load.

An agent can never grant itself more freedom than allowed

This is the promise that makes autonomy safe to hand out. An agent declares the level it wants. Your organization sets a ceiling — the most freedom any agent is allowed under a given team, role, or scope. The agent always ends up at the stricter of the two. It can ask for more; it can never take more.

Ask for the moon, get the ceiling. The organization's limit always holds, so autonomy can only ever tighten — never loosen.

Because the ceiling is checked when an agent is shared more widely, an agent can't quietly gain freedom just by being promoted to a bigger audience. If its requested level is bolder than the ceiling for that audience, sharing it is refused until the mismatch is resolved.

Autonomy lowers friction — it never overrides a hard stop

Autonomy only relaxes the routine "please confirm this write" checkpoint. It sits underneath your explicit guardrails, never on top of them. Some approvals always happen, at every level, including autonomous:

An action a specific approval rule requires — say, "every refund over $500 needs a manager" — always pauses.
An agent explicitly marked always ask first always pauses, whatever its level.
Spending limits, data-loss protection, tool restrictions, and the emergency stop keep working exactly as before — autonomy touches none of them.

Freedom, not a free pass Turning an agent up to autonomous makes it faster on the everyday. It does not switch off a single one of your organization's other protections. The rule you wrote to always require sign-off still fires — every time.

How to think about setting it

Start supervised
Every new agent begins on the safe default. You watch it work, approve its changes, and build confidence with nothing at stake.
Graduate to trusted
Once an agent has proven itself on the routine, let it handle the everyday on its own — while it still stops at the edge of anything it can't take back.
Reserve autonomous for the proven
For a mature, tightly-scoped agent whose other guardrails already carry the risk, remove the last checkpoint and let it run.
Set a ceiling for peace of mind
Cap how far any agent in a team or role can go. Now individual owners can tune freely, knowing the organization's limit can never be crossed.

Why it matters

Move fast where it's safe

Stop rubber-stamping low-risk work. Let proven agents handle the routine at machine speed.

Stay careful where it counts

Keep a human on every irreversible action, automatically, without slowing down everything else.

Govern with confidence

One organization-wide ceiling means no agent can ever exceed the freedom you've decided to allow.