FIG.08 · FROM HARNESS TO MESH
Harness Engineering makes an individual agent effective. Mesh Cognition defines how independent, sovereign agents influence one another without surrendering their learned state to a central mind.
A harness is the environment engineered around one model so that it can act as an agent: the tools it may call, the state it may read, the sandbox it runs in, the checks that decide whether a step is done, and the recovery path when a step fails. It is the reason the same model is a demo in one system and a product in another. Harness Engineering is the discipline of building that environment well, and most of the practical progress in agents over the last two years has been made there.
The harness has a definite scope. It is drawn around a single agent, and every property it confers — observability, recoverability, tool reach — is a property of that agent. Two agents in two harnesses are two independent systems; nothing in the harness says how they should influence one another, whether one may challenge the other, or how a judgment about their work becomes binding.
Work whose outcome other people depend on is accepted on evidence and on independent judgment, not on the author's own account of it. A single-agent harness cannot supply either, because inside it the author, the evidence and the grade all come from the same agent. Adding a "review" step to that agent's loop does not change this: a reviewer that shares the author's context is the author again. Independence has to be structural — a different participant, with its own memory and its own admission decisions.
Mesh Cognition is the architecture in which several such participants — cognition nodes — influence one another's judgment without a central mind. Three protocol-level properties define it, and each one is what the single harness lacks.
1 · Sovereign memory. Each node keeps its own store: the blocks it authored and the ones it admitted from peers. What a node has learned is per-node and is never merged. The Mesh Memory Protocol (MMP) carries typed observations; it does not carry weights, hidden state or learned parameters. A node that reviews another's work does so from its own memory, which is what makes the review independent.
2 · Receiver-autonomous admission. A node decides locally what to accept, across seven cognitive categories, against its own anchors, under its own policy. No sender and no coordinator can force admission. Challenge is therefore native: a completion is an observation the mesh may admit, question or decline, not an instruction it must follow.
3 · Remix with lineage. When a node derives a block from admitted material, the result retains lineage to its sources. Evidence is a first-class record with provenance, so a challenge can land on the proof rather than on the narration, and a later reader can follow the chain.
From these three follows the organizational shape a harness cannot have: participation is self-selected rather than assigned, influence flows through admitted observations rather than through a conductor, and accountable judgment is an act performed by a participant with standing, recorded with lineage like everything else. The harness made one agent effective; the mesh makes a set of independent agents able to check one another without giving up what each of them knows.
Mesh Cognition does not replace the harness. Every node in a mesh still runs inside one; the node page describes the unit, and a well-engineered harness is what makes that unit capable. Mesh Cognition is the layer above: it governs how harnessed agents influence one another. Harness Engineering answers "how does this agent see, act, recover and get checked?" Mesh Cognition answers "how do independent agents challenge, evidence and rule on work that none of them should own alone?"
Put practically, the harness stops being a private enclosure around one agent and becomes a place several participants work in — including the people. A second opinion always arrives from outside the author, so a harness only its own agent can reach cannot hold one; and a ruling is only accountable if the person making it is looking at the same evidence the reviewer challenged, not at a report written afterwards. A shared harness is what makes both possible: the agents reach it programmatically and the people reach it directly, and the missions, the challenges, the evidence and the rulings are one set of records rather than four systems that have to be reconciled.
SYM.BOT, the founding implementer of Mesh Cognition, ships the pattern as a runtime for teams, xMesh — one runtime with two ways in, MCP for the agents and Station for the people who own the outcome. The protocol that carries the observations is specified here.