The missing layer in the agent stack

Ask an engineering team to draw their agent stack and you get the same picture everywhere: a model at the bottom, a harness around it — planning, tool calls, memory — and a growing column of integrations on the side. Two years of harness engineering standardized that machine half of the diagram at remarkable speed.

Now ask the same team what their agent knows about the person it is working for. The answer, almost always, is prose: a resume pasted into a system prompt, a few chat logs, an "about me" paragraph someone wrote in 2023. The machine half of the stack got a discipline. The human half got a copy-paste.

Where the stack actually stands

The layering happened fast because each piece solved a supply problem:

• Models supply reasoning.
• Harnesses supply control flow — planning, retries, tool routing, evals.
• Protocols supply integrations — one tool schema, many clients.
• Memory supplies history.

Every box in that list has a stable schema, a clean transport, and explicit rules about who may read what. Every box, except one.

The blank spot: the person

The input the harness least understands is the one the whole system is nominally serving. Today, human context fails in three specific ways:

1. It is unverifiable. An agent that "knows" you from a pasted resume is guessing from an advertisement. Nothing links the claim to a source, so nothing can be checked.
2. It is unversioned. People change — roles, interests, constraints. Prose in a prompt drifts silently against reality, and nothing tells the harness which version of a person it is talking about.
3. It is all-or-nothing. A pasted document exposes everything in it. There is no way to say "this client may see how I work, but not why" — no scopes, no revocation, no log.

These are not prompting problems. You cannot fix them with a better system prompt, for the same reason you cannot fix an unversioned API by writing nicer docs.

What makes a layer a layer

A layer is something a harness can depend on. That has a concrete meaning:

• a stable schema that does not change shape between clients;
• a transport with one canonical envelope — REST for programs, MCP for agent clients;
• authorization evaluated at read time, not at copy time;
• versions, so a profile can improve without silently drifting;
• revocation, so access ends when the person says so.

Human context has none of these by default. That is the gap.

Our bet: context is a supply problem

ProfileClaw treats career context the way harness engineering treats everything else — as a measurement and supply problem. Four short instruments record how a person actually works: interests (RIASEC), working style (Big Five), values, and demonstrable skills. The results live as a layered, versioned profile that any harness reads through one consent-scoped API.

The practical difference for builders is a shift from inference to reading. Today an agent infers the person from fragments and performs personalization as theater. With a context layer, the harness reads a structured summary — fresh, consented, logged — and spends its prompt on the task instead.

The next frontier in harness engineering is not another tool protocol. Tools already have a story. The missing story is standardized human input — a layer that treats the person with the same engineering respect as the model.