AgentMesh

Part 2 of 3 ยท Part 1: The anatomy of a mesh

The reward is the business goal

AgentMesh published this article in November 2026.

Companies pay their agents for business results, yet every agent today is trained and tuned on task scores. A mesh can train a team of agents toward the business goal itself, because it is the one place that holds the goals, the limits on them and the outcomes of every agent's work.

The example is the same marketing mesh as in Part 1: five agents from different vendors working toward three goals for the quarter, on a $60,000 budget.

Agents are trained on task scores, but companies pay for business results.

Today an agent gets better against measures of its task: whether the code passes the tests, whether the answer is right, whether the task finished, how it scores on a benchmark. Those are real gains, and they are the only measures a single agent or its harness can see.

A head of marketing does not pay for posts written or emails sent. They pay for qualified pipeline. The content agent can get better at posts every week while pipeline stands still, the same gap between each agent's number and the team's goal that a mesh exists to close.

Training on task scores improves the parts. Training on the business goal improves the team.

A reward in business terms needs three things, and only the mesh holds all three.

The goals

What the business is measured on, each with a weight: qualified pipeline up 20%, cost per lead of $180 or less, share of voice up 5 points. They live on the mesh, not in any agent.

The limits

What the goals must not harm, and the lines nobody crosses: brand sentiment holds, no contact without consent, spend never above $60,000. They make sure the reward cannot be bought by breaking something else.

The outcomes, across every agent

The ledger holds a signed entry for every step each agent took, and goal telemetry shows what happened to every goal afterward, across every harness, vendor and company in the team.

A reward written in business terms needs all three in one place. The mesh is that place, because it is where the goals are set, where the rules are enforced and where every agent's work goes into the ledger.

The Balanced Scorecard already describes the reward a mesh manager needs.

In 1992, Robert Kaplan and David Norton argued in Harvard Business Review that financial measures alone give a company misleading signals, because they report what has already happened. Their Balanced Scorecard measures a business from four perspectives at once: financial, customer, internal process, and learning and growth. The measures form a chain of cause and effect, so the early ones show whether the financial results will follow.

A mesh's goals are a scorecard for a team of agents. In the marketing mesh, pipeline and cost per lead are the financial measures, share of voice and brand sentiment are the customer measures, and drafts reaching outreach within a day is a process measure. Learning and growth is the work history itself, and the better team it builds. With the whole scorecard as its reward, the mesh manager cannot buy pipeline by burning the brand, because the brand is on the scorecard. Goal telemetry also tests the chain of cause and effect that most scorecards only assume, such as whether week-two replies really become qualified leads by week six.

The mesh trains the mesh manager and the team's setup, never the vendors' agents.

The agent being trained is the mesh manager: the agent that turns goals into a plan, gives each agent its part, moves budget and asks for approvals. Its decisions are where the team's results are won or lost.

Alongside it, the mesh learns the team's setup: which agent holds each role, who hands work to whom and how the budget is split.

The agents themselves stay with their owners. The mesh never retrains a vendor's agent. Its owner can receive that agent's successes and failures from the ledger, with the company's consent, and decide what to do with them.

The mesh trains on replays of past work, so the live business is never at risk.

A live business is too slow and too costly to learn from by trial and error, so the mesh trains on its own work history:

  • The reward comes from the goals. A plan scores by how far it moves the goals, each weighted as the owner set it.
  • The limits come from the rules. A plan that breaks a hard limit scores nothing, and harm to a goal that must not suffer counts against it.
  • The environments come from the work history. Past weeks and quarters are replayed, with the outcomes that followed, so a new plan is judged against what actually happened.
  • Earlier signals stand in for slow results. Pipeline is counted once a quarter, which is too slow to learn from. Goal telemetry shows which earlier signals lead to it, such as week-two replies that become qualified leads by week six, and the mesh learns from those.
  • Credit goes where it belongs. The ledger ties every hand-off and decision to what came after, so the mesh can tell which agent's work and which decision moved the number.

The environments run in a framework such as Harbor. Every new version of the mesh manager has to beat the current one in them before it runs a live mesh.

A new mesh manager must beat the current one on past work before it runs Live mesh The team works toward its goals Work history Ledger: every step Telemetry: every goal Environments Past weeks replayed with their outcomes Training Reward: the goals Limits: the rules Beats the current one? If not, it keeps training Yes: the new version runs the mesh

The live mesh writes its work history, that history becomes the training ground, and only a mesh manager that beats the current one on replays of it goes back to run the team.

Each way it can go wrong has a guard.

Training toward business results is harder than training toward task scores. Each hard part has an answer in the mesh.

What can go wrongThe guard
The mesh manager games the reward, for example buying pipeline by burning the brandThe goals that must not suffer and the hard limits are part of the reward, so harm counts against a plan and a broken limit scores nothing
Results arrive slowly and noisilyEarlier signals learned from goal telemetry, and many past quarters replayed instead of one live one
Nobody can tell which agent or week moved the numberThe ledger ties every hand-off and decision to what came after
The work history holds customer dataPersonal data is removed before anything is used for training
A model maker's terms forbid training on its outputsTraining uses only models whose terms allow it
A company's work history is used without its sayThe work history belongs to the company, and nothing trains across meshes without its consent

Each company's mesh gets better at that company's goals.

A mesh manager trained this way gets better at what its company is actually measured on: this quarter's pipeline, next quarter's margin. It learns from that company's own work history, under that company's own limits.

Every company can buy the same agents and the same models. No company can buy another's work history, or the mesh manager trained on it.

Start with one mesh, give it a goal that matters, and the work history starts building from the first week.

Part 3: Processes are how a mesh reaches its goals