Open
Each agent reasons on its own Kernel-granted slice. No agent sees another's raw context — conclusions publish to a shared, governed board.
ECHO MULTI-AGENT LAYER · COLLABORATION · ON THE LOCKED CONSENT MODEL
Agents reason together. The Kernel brokers every hop.
This is the layer above the Consent Model: multiple agents actually build on each other across rounds. But no agent ever hands context to another. Every cross-agent hand-off is a Kernel-brokered Context Grant — minimized, purpose-bound, logged. An agent's conclusions re-enter the Kernel as new governed context before any peer can use them, and the final decision is the human's alone.
The shared boardThe debate's working memory is Kernel-held, never agent-held. Agents stay pure functions: a scoped slice goes in, a proposal comes out, nothing is retained. What compounds across rounds lives in the Kernel — governed, attributed, and auditable.
Personal memory is locked. Samples remain available.
Three rounds, one boundary
Cross-examine
Agents build on peers' claims. Each claim re-enters the Kernel and is re-minimized before a peer receives it — never handed over directly.
Converge
Final positions on the shared board. A synthesis is proposed — never decided. The decision returns to the human.
LAW 4 · AMENDED
No agent may pass memory, instructions, or derived claims directly to another agent except through Kernel-approved context. Collaboration is where that clause earns its keep: a conclusion is new context, and new context is governed like any other.
Where it sits
vs. /consent
The Consent Model is the broker: it issues one Context Grant and enforces the tiers. Collaboration runs that broker many times, across rounds, as agents build on each other.
vs. /council
The Council deliberates once — four scoped views, one pass. Collaboration is a real debate: claims compound round over round, each hand-off re-brokered and logged.
vs. /agent-mesh
The Mesh proved a single minimized disclosure between two agents. Collaboration closes the dangers the Mesh deferred: transitive re-sharing and agent-to-agent instruction passing.
⚠
Honest sharp edges
Collaboration is the second step of the design order — consent model → collaboration → unification. These are the places where discipline, not a checkbox, keeps a multi-round debate from becoming a side channel.
Compounding is where laundering hides
Each round adds context. The safety is that a derived claim is never raw memory and never travels directly — it is re-minimized on every hop. That discipline must hold as reasoning deepens.
Consensus is not truth
Agreement among agents is a signal, not a verdict. The synthesis is a proposal; the human still decides, and can reject a unanimous board.
Board classification stays conservative
Derived claims enter as low-sensitivity summaries, labelled inferred and attributed. If a claim would reveal a high-tier fact, it is gated like any other high-tier context.
Session boundary still bites mid-debate
A high-tier cross-examination needs an active consent session. When the session ends — explicit close or idle timeout — those grants expire even if the debate is unfinished.
Debate is cheap; the boundary is everything. Agents can think together for as long as the problem needs — and still, not one of them can act, know beyond its grant, or decide what you meant to keep.