Most AI projects stall in the same place: the demo works, then someone asks what happens when it's wrong, and nobody has a good answer. The model is rarely the problem. The operating discipline around it is.
At TRW Labs we run our own operations on a staff of AI agents: one keeps the books on deadlines and renewals, one watches security every night, one handles the inbox, one writes and posts here, one reviews everything before it ships, and a chief of staff ties it together into a single morning brief. Building that taught us more than any benchmark. These are the five rules we hold every agent to, for ourselves and for clients.
1. A person stays in command
Agents draft. People decide anything that leaves the building: sending an email, publishing a post, spending money, deleting something. That sounds slow until you make approval fast. Every decision in our morning brief arrives with a link; one tap approves it, one tap declines it, and the agent does the rest. Command doesn't have to mean busywork.
The test is simple: if an agent can do something that would embarrass you, cost you money, or can't be undone, a human answers first.
2. Put the limits in code, not in the prompt
Instructions are suggestions. Permissions are rules. If an agent must never send an email without a yes, don't write that in its instructions and hope. Take the send button away and give it a drafts folder.
Every agent on our staff has its own access list. The security agent can look at everything and change nothing. The writer can post to one platform and no other. What an agent can't touch, it can't break, and no model gets to decide on its own whether an action is safe.
3. Write everything down
An agent that forgets is an agent you have to re-brief, and an agent whose memory you can't read is one you can't correct. Our agents keep their memory in plain files: what they know, what's open, what they decided and why. Every run ends with a line in a shared log. If a conversation disappears, nothing important disappears with it.
The same habit makes audits possible. When something looks off, you read what happened instead of guessing.
4. Give every agent a number
"Helps with marketing" is not a job. "Five qualified leads a week, each with a source and two lines on why it fits" is. Each agent gets weekly targets you can count, and once a month every one of them is audited against its record: keep it, rewrite its job, or retire it.
The real risk with AI staff isn't that it fails loudly. It's sprawl: agents that run quietly, produce nothing anyone uses, and still cost attention. A silent run and a successful run look identical from the outside unless something is counting.
5. Measure the cost and the payoff
Every agent should earn its place in one of three ways: hours given back, money moved, or risk caught early. Anything else is overhead, however clever it looks. We track usage every day and ask one question every Friday: what did the staff actually save this week?
If the answer is "not much", the staff gets smaller. That's a feature, not a failure.
The model is the easy part
Capable models are everywhere now. What separates an AI system you can trust with real work from a demo is the structure around it: who's in command, what's enforced, what's written down, what's measured. That's the work we do at TRW Labs, for our own operations and for clients who want AI that runs real work without anyone losing control.