Agents as a workforce

The problem

Most teams add agents the way they add tools. Someone finds one that works, a few people start using it, and six months later nobody can say what it owns, who checks it, or what it is never allowed to touch.

That is not a workforce. It is a pile of helpers. The first time one of them ships a wrong claim, the team learns to distrust all of them.

The operating model comes first, and it is a process. Stages. One artifact at each handoff. One gate. Messaging split into what is locked, what can be adapted, and what is free to write. Casting that process as an agentic system does not skip the process. An agent with no stage to run is just a faster way to improvise.

The point of an agent is to take repetitive work, so people spend their time on strategy, judgment, customers, and creative. Agents add a layer. They do not replace the team, and they do not replace the lead model already running.

The system

Build the process, then put agents on it, then deploy them in an order that can be stopped.

1. 1Process
2. 2Memory
3. 3Agents
4. 4Evaluation

The process is the foundation

A campaign already has a path. A person writes the strategic brief and a person signs it. The brief has an objective, an audience, messaging, success metrics, a timeline, and a budget. Then adaptation. Then build and produce. Then launch and execute. Then report and learn. Each stage ends in a gate, and each gate has an artifact. No artifact, no handoff. That rule applies to an agent the same way it applies to a person.

1. BriefA person writes it and signs it.
2. AdaptInside the tier.
3. BuildProduce the work.
4. LaunchExecute.
5. LearnThe report, then the next brief.

Messaging has three tiers. Locked messaging is used as written. The agent may choose what to emphasize. It may not rewrite it. Flexed messaging can be adapted inside the argument, and a person approves it before it ships. Free messaging can be written for a local follow-up, inside the guidelines, and the draft is logged.

1. LockedUsed as written. The agent may emphasize it, not rewrite it.
2. FlexedAdapt the argument. A person approves before it ships.
3. FreeA local follow-up, inside the guidelines, and logged.

If the stage, the artifact, and the tier are not written down, do not put an agent on the work. You would be automating a meeting that has not been designed yet.

Five lanes

Five agents cover the campaign. Each one gets a lane in one sentence, and a person who approves it. If either is missing, it does not go live.

1. 1Intake
2. 2Decomposition
3. 3Steward
4. 4Rollup
5. 5Governance

Intake takes a finished brief, or a ranked piece of work, and nothing else. It does not turn every signal into a task. A customer-confirmed need gets help. A verified change gets a short look. A single inferred visit gets no agent task at all. High-fit accounts with no signal still get a person’s attention, because waiting on a score is the dependency you were trying to leave.

1. ConfirmedThe customer confirmed the need. Give help.
2. VerifiedA real change. A short look, not a campaign.
3. InferredOne visit. No agent task.

Decomposition reads that brief and breaks it into tasks. It drafts inside the tier it was given. It does not invent facts, and it does not open a new value proposition.

Steward keeps routine facts current. Status. Dates. What moved. It may write those updates on its own only after it has a record of being right. Until then, a person checks them.

Rollup writes the report of what changed and puts the learning where the next brief can use it. A person reads the report and writes that next brief. The loop does not close by itself.

Governance sits on the gates. It is the lane that stops the others. It checks the tier, the citation, and the never-do list. It does not ship the work. A person does.

The path is short. A person signs the brief. Intake accepts it. Decomposition drafts the tasks. Governance holds anything that is not allowed. A person approves. The work runs. Steward keeps the routine record current. Rollup brings the learning back. A person writes the next brief.

Build, then deploy

I would build in four moves, and I would not swap them.

Process first, as above. Then memory. One identity for the account across the sources. A schema written down before anything is loaded. A citation on every fact, so a later agent can tell a customer-stated fact from a sales report, an approved line, a figure from analytics, an inference, and a fact that is old or unverified. Then the five agents, on that process and that memory. Evaluation last, and it is not optional. Without it you can see that the system is running. You cannot see that it is working.

Two more rules sit inside that order. Who can see a fact is decided before an agent is allowed to read widely. And writeback is earned. Routine steward updates can write themselves once they are proven. Claims, pricing, positioning, and roadmap facts go through a person, with no end date on that rule. The place the team already works, and any signal from outside the company, come after this core is solid. A new interface on top of an uncited store gives you confident wrong answers, faster.

Deploy in the same spirit. Start on-demand, the kind that runs when a person asks. Let an always-on agent earn that status. A lane that only watches may flag. It does not fix, block, or decide until it has a record that justifies more.

• On demandRuns when a person asks. This is where a lane starts.
• Always onEarns it. A watcher may flag. It does not decide.

If the review pile grows faster than the people who approve it, narrow what the agents are allowed to start. Do not add agents to outrun the approvers. The specialists are the constraint.

What stops an agent

Four stops, and governance is what enforces them.

1. 1Tier
2. 2Citation
3. 3Never-do
4. 4The record

The tier. Locked copy is not rewritten. Flexed copy waits for a person. An agent does not introduce a value proposition that was not in the brief.

The citation. Every fact carries where it came from, and a date. An inferred fact is not a customer-stated fact. An outdated fact is not settled. Public information is labeled public.

• StatedA customer said it, or a person approved it.
• DerivedIt came from a report or from analytics. Say so.
• InferredThe agent concluded it. It is not a fact yet.
• StaleOutdated or unverified. Do not ship it as settled.

The never-do list. Claims, pricing, positioning, and roadmap facts stay with a person. So does creating the work in the system before someone says yes. So does treating a single inferred signal as a reason to open a task.

The record. Nothing on the always-on side, and no unattended write, until the lane has been right often enough that a person will sign for the pattern.

What still belongs to a person

Strategy. Whether a claim is true. The customer conversation. The creative idea. The call on what to stop funding. The review of what the agent drafted. The first offer that has to be useful, with a defined output and a cap on expert time. An agent can draft that offer. A person is accountable for it.

• A personStrategy, the claim, the customer, the idea, what to stop, and the approval.
• An agentThe draft, the routine record, and the report. The offer still has an owner.

Approval is a skill. The person catching a wrong claim, a stale number, or a promise the product cannot keep is doing the job. A team that treats that review as a rubber stamp has agents without oversight.

The numbers

I am not tying a result to this system. Evaluation is the last move on purpose, and it is not measured yet. The pipeline-per-program-dollar result has its own guide, and I am not borrowing it here.

When evaluation exists, four watches belong on it. Whether the answer was accurate. Whether the citation was the right kind. How often a fact was stale. And time returned to strategy, which is the point of taking the repetitive work off people. There is no baseline and no period for any of them yet.

• AccuracyWas the answer right. Not measured yet.
• CitationWas the source the right kind. Not measured yet.
• Stale factsHow often a fact was old. Not measured yet.
• Time returnedRepetitive work handed back to strategy. Not measured yet.

What you can take with you

The failure mode is an unmarked fact. An inference, or an outdated line, gets treated as settled because nothing said where it came from. A person catches it at the end. That is the expensive place to catch it. Citation exists so the catch happens at the gate, while the draft is still a draft.

What I would change is the order. Write the operating model before the first prompt. Put the five lanes on it, each with an approver and a line on the never-do list. Build the schema and the citation before an agent is allowed to write. If the review pile grows, cut the scope of what the agents start.

1. 1Operating model
2. 2Five lanes
3. 3Schema and citation
4. 4Cut the scope

Here is the checklist.

• Write the operating model first. An agent runs a stage that already has an artifact and a gate.
• Give each agent one lane: intake, decomposition, steward, rollup, or governance. Name who approves it.
• Start on-demand. A steward writes routine updates only after it has a record.
• Keep claims, pricing, positioning, and roadmap facts with a person, with no end date.
• Mark where every fact came from. An inferred fact is not a customer-stated fact.

An agent that runs a stage you already designed, and hands the work back with an artifact and an approver, is a good colleague. An agent that quietly ships its own work is a liability, however good the draft looks.