The Uderia Agent
an anatomy of trust

The one idea to keep

Brain and senses are constitutional. Instruments are acquired.

The revision turns on a single distinction that your original sketch already had right: the MCP data source is not a tool the agent picks up — it is a sense the agent is born with.

The five faculties

Five points, one identity

The star stays — it's the right hero for an agent-centric platform. Each point is a distinct faculty: what the agent is, what it perceives, what it remembers, what it knows, and what it can do. Together they make one identity — and the brain that reasons over them is born into the DNA, not bolted on. And no agent is truly alone: the faint constellation around the membrane is Coordination, one agent composing many.

The two foundations

What the star stands on

Grounding and Guardrails are the load-bearing floor — the agent's principles, not its parts. One keeps it honest about what it knows; the other keeps it safe in what it does.

Grounding's role is Truth / Faithfulness, not Values — "values" reads as ethics, which is what Guardrails carry. Grounding is about factual fidelity: the sharpest, most defensible claim the platform makes. And both foundations are graduated, not binary — each runs on the same ladder: off → observe (measure, never alter — the default) → annotate → review (withhold, but keep the original for a human) → enforce. You climb it deliberately, per agent.

The boundary

Trust is compositional, not just a container

Your dashed membrane is the right perimeter — but its most novel property isn't that it encloses the agent. It's that a trusted agent is made of trusted ingredients: the identity is 🟢 only when its ontology, its skills, its configuration each verify. A consumer reads one badge; the composition is what earns it.

So the membrane isn't just a box drawn around the star. Each point carries its own trust state, and the agent's trust is their sum — signable, cryptographically verifiable, and portable across instances. That is the claim the diagram should make out loud, because it's the one competitors can't casually copy.

The membrane breathes

Trust is a closed loop, not a stamp

A signature proves an agent is unchanged. It cannot prove it is behaving. The end-to-end approach closes that gap: measured conduct changes trust, trust can change enforcement, and every step is visible and verifiable. The membrane is not stamped once — it breathes.

Trust in Uderia is a composite of four axes. The star is the first; the membrane, the second; and two more make the picture end-to-end:

These four turn on one turn of the loop. A user request is screened (inbound guardrail), the acting principal is set (authority), the trusted scope is stamped (grounding), every tool call is enforced, the answer is guarded and grounded on the way out, the turn is scored, and the whole thing is sealed. What that loop measures then feeds back: an agent that drifts is caught by evaluation; sustained bad conduct trips a circuit breaker that can freeze its unattended use; a risky answer can be held for a human instead of discarded; and no path is silent — a poisoned memory, document, or tool result is defused at the door, and every incident becomes a worked item in an Assurance Inbox. Before you tighten any of it, you can backtest the policy over real history or red-team the agent — deciding on evidence, not faith.

Changelog

What changed from the first sketch