How To Brief an AI Agent So It Runs the Work Without You
A clear brief lets the agent run the production and QA while the final judgment stays with you.
By Gabriel Tan | July 2026
Your team finished the work hours ago. You are still at your desk because none of it moves until you have read it. You run a communications agency, and everything the team produced today, the client update, the award entries, the talking points, sits in a stack waiting on the one thing only you give: your judgment. You are the queue, and nothing moves faster than you can clear it.
There is a cleaner way to run that. You take the task, you give your comments, the edits happen there and then, and the work runs its own quality checks before it comes back to you. You still own every judgment. You just stop being the bottleneck that holds it.
The shift is in how you hand the work over. You are not writing prompts for an artificial intelligence (AI) tool. You are briefing an agent the way you would brief a sharp intern: give it the source material and the context first, then the task. Do that well and the work stops queuing at your desk.
The method has three parts, and you own two of them
Treat the agent as a very smart intern in their first week. You do not write out the brief in full. You talk, and it writes the brief back to you so you can correct it. Do it by voice if you can. Speaking is faster than typing, and the agent strips out the filler.
The first part is the brief, and you are not finished until it is agreed. You describe two things out loud: the situation, and the outcome you want. Describe the outcome, not the wording. The words are the agent's job. Then you make it work for the brief. It asks you questions, it challenges the weak spots, and it repeats the whole brief back to you. You confirm what it got right. Then you ask it for two more questions you had not thought of. A job with many parts always hides gaps you did not see. You answer them now, or you fix a wrong draft later. Do not let it build until the brief is one you both agree on.
The second part is the build, and it runs in one shot. One instruction: produce the whole thing at once. It builds the full job together and holds it as one piece, instead of stopping after every section to wait for your nod. Then you give the instruction that most people never think to give. Run a four-round quality assurance (QA) pass on your own work before you show me anything. Fact-check first, every claim, number and name against the source files, so nothing is invented. Then narrative, so the piece flows and leads somewhere. Then adversarial, where it attacks its own work and hunts for the weak and the wrong. Then the client lens, where it reads the whole thing the way your client will read it. It does the work, and it checks its own work, before any of it reaches you.
The third part is the review, and this part is yours. You come in, you read, and you give your comments on what is not good enough yet. If it is still short of your bar, you comment again on the exact areas that need work, and that feedback becomes the brief for the next pass. You ask it to report back and repeat your comments so you know it understood them. Then it produces again and runs the four rounds again. You hold that loop until the work meets your standard. The judgment never leaves your hands.
This is the same shape we use to build every workflow for a client. A human frames the work, the machine produces it, a human signs it off. The agent never decides what should be produced, and nothing reaches your client without your read.
What to do this week
Pick one job and run the whole loop once. Choose something you keep pushing to tomorrow.
1. Choose a batch you dread. The emails, the award entries, the briefing notes. Five minutes to pick.
2. Talk it through until the brief is agreed, then give one instruction: build it in one shot, and run the four-round QA on your own work before you show me. Fifteen minutes to brief, then leave the room.
3. Come back and read. Give your comments, then ask it to repeat them back to you, raise any questions it has, and confirm the revision with you before it touches the work. Fifteen minutes.
4. Once the revision is confirmed, it produces again in one shot and runs the four rounds again. Repeat until it clears your bar, usually one or two rounds at ten minutes each. Then it ships, with your name on it, because you were the one who judged it.
That is well under an hour of your attention for a batch that used to take a morning.
Where I was wrong about running agents in parallel
Last week I argued against running AI agents in parallel. I want to correct part of that.
The mistake was treating parallel as one thing. It is two. Production can run in parallel. Judgment cannot, and feedback cannot. You can have five agents producing five different jobs at the same time. What you cannot do is brief two of them at once, or split one round of feedback across three. So you brief one job in full, then brief the next. You review one job in full, then review the next. The agents run side by side on the production. Your judgment goes one at a time, start to finish, and so does your feedback.
Hold that line and the numbers move. Three jobs at once becomes five, then more, with the quality holding, because you never once multitasked the part that only you can do. The moment you try to give feedback on two jobs in the same breath, both get worse.
Go back to the pile on your desk. The change was never that you type faster. It is that you stop being the place where work goes to sit. You describe the job clearly, you agree the brief, and you trust a checked process enough to leave the room. The question stops being how much you can clear tonight. It becomes how much you are willing to hand over.
Gabriel Tan is the founder of Mekong Bridge Advisory. He builds structured execution systems for PR and communications firms.
