Part 1 of 3
The anatomy of a mesh
AgentMesh published this article in November 2026.
A mesh is everything around a company's agents that turns them into a team working toward goals. Every company already runs agents in many harnesses: coding agents such as Claude Code and Codex, managed agents from Anthropic and the other model makers, and agents built into the software it pays for. Each harness makes its own agents useful, and no harness turns all of them into one team.
The agents do the work. The mesh is what they work inside: the goals they share, the rules they follow, and a work history of everything they did, which is how the team gets better.
The example throughout is a marketing mesh: five agents from different vendors working toward three goals for the quarter, on a $60,000 budget.
A mesh wraps the agents a company already has in shared goals, rules and a work history.
Agents and the harnesses they run in sit in the middle. Everything around them is the mesh, and goals are its most foundational part.
Agents become a team only when they work together for a higher goal.
Agents become a team only when they work together for a goal higher than any one of their tasks. The content agent writes posts, the outreach agent sends emails and the ads agent buys clicks. Each can hit its own number while qualified pipeline, the goal marketing is measured on, stands still. A mesh gives them that higher goal and keeps every agent pulling toward it.
That goal belongs to no single agent and no single harness. A company's agents run in many harnesses, and each harness sees only its own agents, so the mesh has to do what no harness can:
- Hold the shared goal, and weigh it against the other goals the team must balance.
- Bring in every agent that can help, whichever harness, vendor or company it comes from.
- Keep all of them inside one set of rules and one budget.
- Pass work between them, and keep a record of all of it.
- Settle it when one team's goal pulls against another's.
Each of those is a part of the mesh.
A mesh has ten parts, and the team falls short without any one of them.
Each part gives the team one behavior it needs.
Every agent works toward the same goals, weighed against each other.
Goals are the most foundational mesh primitive. A mesh holds several goals at once, each with a weight, the tradeoffs between them priced, what they must not harm, the hard limits and one budget for the whole team. The parts of a goal come from goal engineering; a mesh holds them for a team instead of one agent. Without goals on the mesh, every agent optimizes its own number, and the team can hit every target while the business number stands still.
Any agent can join from any vendor or harness, and everyone knows who owns it and what it may do.
The roster lists every agent in the mesh, who owns it and what it may do. Each agent, harness and mesh describes itself once, in a descriptor, and publishes a card made from it, so any member can check who it is dealing with before it acts. Without a roster, nobody can answer the first question an auditor asks: which agents work for this team?
Agents that speak different protocols work together without changing.
Agents do not share a language. Each harness and each vendor speaks its own protocol, and no company will rebuild its agents to match the others. The mesh mediates: every agent joins in the language it already speaks, and the mesh carries the meaning across, so any agent in the team can work with any other. That is what connecting agents means in AgentMesh. Without mediation, a team can only include agents that already speak the same protocol.
No agent crosses a line, whoever built it.
The rules belong to the mesh, so they bind every member, whoever built it and wherever it runs. Spending stops at the budget, a decision reserved for a person waits for that person, and any agent can be stopped in seconds. Without them, each agent's limits are whatever its vendor shipped.
Work outlives a session, and any agent can pick it up.
Rooms for working together, a shared drive, memory, feeds and a work board hold the team's state outside any one harness. An agent that joins on Tuesday can read what the others did on Monday. Without them, work lives in sessions and ends with them.
Agents pass work to each other without a person carrying it.
Agents pass work to each other directly, agent to agent, and use the mesh's own services, agent to mesh. Every hand-off is addressed by name, delivered and written to the ledger, whichever transport carries it. Without hand-offs, a person carries every result from one tool to the next.
Someone plans the work and changes the plan as results come in.
The mesh manager is an agent that runs the mesh. It turns the goals into a plan, gives each agent its part, watches the results and changes the approach inside the rules, and it asks a person when a decision is theirs. Without it, the plan lives in someone's head and changes only when they have time.
Anyone can see what happened and whether the goals moved.
Every hand-off, action and approval leaves a signed entry in the ledger: a receipt of who did what, when and at what cost. Goal telemetry reads every goal as results come in. Together they answer two questions at any moment: what happened, and is it working? They make up the team's work history across every harness, and the team learns from it. Without them, the team cannot learn and nobody can audit it.
The team learns from its own work history and gets better over the quarter.
Over a quarter the mesh changes more than any single agent can: which agents it holds, which agent hands work to which, how the budget is split and how much each goal counts. It can borrow an agent from another mesh for a while. Each change comes from what the work history shows, and it must pass the team's checks before it sticks. Without this, the team that starts the quarter is the team that ends it, however it performs.
Two teams' goals stop fighting each other.
A company runs many meshes, and their goals pull against each other: marketing wants more spend, and finance wants lower cost. A company mesh above them holds the goals that settle the conflict, and each team mesh balances its own goals inside them. Without it, the loudest team wins.
A mesh learns from its team's work history.
The ledger shows every hand-off, action and approval, and goal telemetry shows what happened to each goal afterward. Together they show where the team wins and where it loses time, money or ground on a goal. The mesh reads them every week and turns what it finds into changes to the team.
Only the mesh holds the whole work history. Each harness keeps a log of its own sessions, but the question "why did pipeline stall in week 3?" needs the hand-offs between harnesses, the approvals, the budget moves and the goal readings together. Those all live in the mesh's ledger and goal telemetry.
Start with one mesh and the agents you have, so the work history starts building. Then the mesh runs this loop every week.
Reading the work history shows where the team loses: work waiting at a hand-off, an agent whose output no goal uses, a rule hit again and again, an approval that always says yes. Each of those becomes a check before the team changes.
The mesh learns at three levels:
The mesh tunes who is on the team, the hand-offs, the budget, the goal weights, each agent's instructions and the shared memory. Every change is tested against checks built from past work. No model's weights change, and it works with every agent, including managed agents from the model makers.
The mesh manager is trained with reinforcement learning in environments built from the work history: replayed situations with automated checks that score whether a plan met the goal inside the rules. The environments run in Harbor. Every new version of the mesh manager passes them before it runs a live mesh.
Each agent's owner receives that agent's successes and failures from the ledger as training data. The mesh never trains an agent it does not own.
The work history belongs to the company that runs the mesh. Nothing leaves it, and nothing is used across meshes, without its consent.
In the marketing quarter, the ledger showed content's drafts waiting at SEO for two days. The check became "drafts reach outreach within a day". The change that passed it was to route content straight to outreach and retire the SEO agent.
The open problems are about teams that are larger, wider and quicker to change.
Some parts of today's harnesses will move into the models, and some mesh parts will move into agents and harnesses the same way. The mesh keeps the parts no single agent or harness can hold: the team's goals, its rules and its work history.
- Pricing the tradeoffs. Today a person writes what one more lead is worth in spend. A mesh should learn those prices from its own results and propose them for approval.
- Meshes across companies. An agent from a partner's mesh should be able to visit yours on a short pass, under your rules, with every step it takes in your ledger.
- Agents built on demand. When a goal needs an agent that no vendor sells, the mesh should be able to have one built, checked and added to the roster.
- Learning across meshes. Work histories read across many meshes should let the next marketing mesh start ahead.
- Hundreds of agents. One mesh manager can direct five agents. Directing hundreds needs meshes inside meshes, each with its own manager.
The work history a mesh keeps is the asset that compounds.
A team of agents writes down everything it does. In a mesh, that work history is the most valuable thing the team produces. It grows faster than any person could read, so the mesh reads it with small models trained for the job.
- Every failure becomes a test. The checks pile up into a test bench for the team, and every change to the team has to beat it.
- Every role finds its best fit. The mesh tries agents, models and instructions for each role against that bench, and keeps whatever does more for the goal.
- Proven work gets cheaper. Once the ledger shows a role done well, that role can move to a smaller open model fine-tuned on the same work, within each provider's terms.
- The team keeps learning for years. Lessons flow back into instructions, into who is on the team and, for the agents a company owns, into the models themselves.
None of that starts until the work history does, and the whole work history exists only where the whole team works: in a mesh. Every company can buy the same agents. The work history its mesh keeps is its own, and it compounds every week.
Start with one mesh, where agents already do real work, and give it a goal that matters.
Part 2: The reward is the business goal