Internet of Cognition¶
Problem statement
Agents can exchange messages reliably. They do not yet reach a shared interpretation of the work they exchange messages about.
Several capable agents assigned to one task will exchange messages correctly and still produce an outcome worse than any of them would produce alone. They commit contradictory findings to shared state, interpret the same intermediate result differently, and terminate a session without a validated agreement. None of this surfaces as an error: there is no exception, no latency deviation and no failed call.
Left unaddressed, these failures degrade the quality and correctness of the result while remaining invisible to conventional monitoring. At that point the premise of multi-agent collaboration — that a group of agents produces a better outcome than any one of them alone — can no longer be relied upon.
Why this happens, and what it costs
Required capabilities¶
Three capabilities are required that a message transport does not provide. Agent systems currently implement none of them.
-
Shared intent
Agreement on what is being solved, in whose terms, and what constitutes completion, established before the work starts rather than inferred afterwards.
-
Shared context
Memory that persists beyond the session, so that results established in one interaction are available in the next rather than being discarded.
-
Shared reasoning
Conclusions that agents build on rather than restate, so that a decision reflects the evidence held across the group.
Architectural principles¶
Five principles shape the architecture. Each addresses one of the capabilities described above.
Convergence rather than delivery. An agent transport — A2A or MCP, carried over SLIM — succeeds when a message arrives intact. The Cognition Protocols operate above the transport under a stronger success condition: every participant ends the exchange holding the same interpretation of what is being solved, in whose terms, and what was agreed. The unit of exchange is a fragment of an agent's cognitive state rather than a payload.
Insights rather than additional telemetry. An agent's telemetry records what that agent did. Collective behaviour is a property of the relationships between agents, so no single agent's telemetry contains it. The Cognition Fabric therefore recomposes the interaction first: one execution model, one trajectory under a single session identity, from which it builds a knowledge graph of agents, tasks and concepts. The analysis runs on that graph, never directly on the raw telemetry spans, and its output is a set of Insights: consolidated findings about the collective behaviour, each traced back to the event that produced it. A named cognitive failure is a good example of such insights.
Understanding accumulates. Findings are written back, so that one engine's conclusion becomes another's input and each submission arrives enriched by what is already established. Without this persistence the system is episodic: each session begins from an empty state.
Capability is granted by tier. The fabric exposes exactly three surfaces, in increasing order of intrusiveness — Analytics (read-only), Active (interact with agents and tools) and Enforcing (block, modify, inject). What an engine can do is determined entirely by which tiers it binds, and only Enforcing needs a component in the traffic path.
Detection closes a loop. A detected cognitive failure becomes an event that re-opens alignment among the agents while the session is still running. Three loops operate over the same fabric: the Cognition Protocol control loop, engine-to-agent interaction, and enforcement.
The architecture¶
Greyed: the Cognition Fabric's API, management and sync planes, the Cognition Engines, and the enforcement and interaction loops — designed, not yet shipped.
The diagram reads from the bottom. Agents operate on an A2A / SLIM / MCP substrate, which the AgentBridge joins as a declared endpoint. Cognition Protocols operate above it to establish shared intent and reach decisions. The Cognition Fabric holds memory and contracts, with SCOPE and the Platform Engines within it converting telemetry into named cognitive failures. Cognition Engines form the judgment tier above. MAS-Lab spans the stack, since any component can be run and measured within it.
Every component, and how they compose
Components documented here¶
Three components have their own sections. The Cognition Fabric and the Cognition Engines appear in the architecture above and are described under Architecture; they do not yet have dedicated sections.
-
Cognition Protocols
The protocol family that establishes and validates shared cognitive state. ACCORD establishes shared intent by coverage; CONCORD selects an outcome under the conflict that remains.
-
SCOPE
Supplies the comprehension layer: ingestion, normalization, the multi-layer knowledge graph, and the analysis engines that convert traces into insight.
-
MAS-Lab
The validation environment. Repeatable, observable and governed configurations in which protocol claims are measured against a matched baseline.
Adoption¶
Adoption is incremental rather than a migration. Three cumulative consumption modes are defined. The first adds the protocol skills to existing agents, optionally through the AgentBridge; the second adds SCOPE and the Platform Engine, which closes the loop by firing the protocol response for the failure it detects; the third runs the system under MAS-Lab for benchmarking and validation.
How to read this site¶
| Topic | Start at |
|---|---|
| Why the project exists | Motivation |
| How the pieces fit together | Overview → Architecture |
| What agents actually exchange | Cognition Protocols |
| How the system sees what happened | SCOPE |
| How the claims are measured | MAS-Lab |
| How to adopt it in an existing MAS | Getting Started |
Tutorials & Demo¶
In preparation
Each component section carries its own Tutorials and Demos page. Walkthroughs are being written; the MAS-Lab quickstart and SCOPE instrumentation guide are available today.
| Component | Tutorials & Demo |
|---|---|
| Cognition Protocols | Demos |
| SCOPE | Getting Started |
| MAS-Lab | Quickstart and tutorials |
| Adoption path across components | Getting Started |
Related reading¶
- Internet of Cognition — the vision, and the Scaling Out Superintelligence whitepaper
- Silent Failures, Loud Consequences — the empirical study behind the problem statement.