# Agents across universes

*Vercy for AI agents · lesson 4 of 4 · ~12 min*

## What you will learn

How agents operate when knowledge crosses sovereignty boundaries: contracts, trust vectors, explainability, and the safety rails that make autonomous semantic work acceptable.

## The boundary changes everything

Inside one model, the register tells you who masters what. Across universes there is no shared register: there are two sovereigns, a negotiated contract, and a protocol. An agent working across the boundary inherits new duties:

- **Consume through Projections, under Contracts.** Never reach for a counterparty's "raw rows" even if technically exposed; the Contract defines what you may take, for what purpose, revocable how. Schema discoverability is not data access.
- **Respect Context.** Knowledge arrives with interpretation context (whose frame, which time, what assumptions). Stripping context to "simplify" is semantic vandalism: the answer you compute will be confidently wrong somewhere subtle.
- **Weigh by Trust Vector, per dimension.** Accept the partner's catalog (high semantic quality) while discounting their inventory numbers (chronically stale). One scalar "trusted" flag cannot express this; do not flatten it.
- **Preserve conflicts.** When two universes disagree, report both claims with provenance and trust: choosing a winner silently is the one sin the whole standard exists to prevent.

## Events make agents auditable

Everything an agent does across a boundary should land as Events: what was requested, under which contract, what was received, what was derived. This is not bureaucracy; it is what makes agent work *disputable*. When a downstream decision is questioned months later, the timeline answers: which projection, which version, which mapping, which confidence. An agent whose work cannot be reconstructed is a liability regardless of its accuracy.

## Explainable results

The standard expects agents to produce **traceable results**: an answer cites the objects, projections, mappings and requirement IDs it used. In practice: prefer "X, per objects A and B under contract C (mapping confidence 0.9)" over a bare "X". Cheap for the agent, transformative for the humans who must act on the answer.

## Safety rails, summarized

Across both courses and the operations guide, the complete refusal set for a conformant agent:

1. Unknown mastership → no write.
2. Mirror, capture, or artifact → no in-place edit, ever.
3. Bypassing classification or the walk → refuse and name it.
4. Contract or conflict rule forbids → refuse, cite the rule.
5. Cross-universe data without a contract → do not fetch, propose federation instead.
6. Cannot attach provenance → do not assert.

Six lines that make "let the agent work on our knowledge overnight" a reasonable sentence.

## Where this is heading

The ecosystem around the standard is building exactly this world: professional agents whose knowledge ships as versioned packs, hired across a marketplace, executing under machine-readable contracts, with track records instead of vibes. The standard's role is the semantic floor: whatever the agent economy becomes, meaning stays owned, history stays honest, and understanding stays verifiable.

## You are done, and just starting

You can now read any conformant model cold, write to it without corrupting it, consume its machine surface, and behave across boundaries. The best next steps: run the cold-start protocol on a real model; read the [Agent operations guide](/spec/#07-guides/Agent-Operations.md) once end-to-end; and when you build your own agent, wire the refusal set in before the capabilities.

## Go deeper

- [AI agent guide](/spec/#07-guides/AI-Agent-Guide.md) · [AI integration patterns](/spec/#07-guides/AI-Integration-Patterns.md)
- [Semantic contracts](/spec/#04-core-concepts/Contract.md) · [Trust model](/spec/#03-federation/Trust-Model.md)
- [Conflict resolution](/spec/#03-federation/Conflict-Resolution.md)