MAXIM
Agent Mesh & Cooperative Intelligence
Sovereign Agents, Cooperative Network
Contents
Design Philosophy
Every Maxim instance is a sovereign agent. It owns its hippocampus, its causal models (NAc), its semantic concepts (ATL), its learned significance weights, and its tool registry. No other agent can modify its state directly.
Cooperation happens through four verbs:
Sharing
"Here's what I learned" — serialized CausalLinks, reflections, or motor programs sent as gifts. The receiver decides what to keep.
Requesting
"Can you do this?" — delegated goals with results returned. The receiver evaluates with its own planner.
Querying
"What do you know about X?" — read-only queries against peer memory. Never returns raw dumps.
Advertising
"Here's what I can do" — capability broadcasts via heartbeat. Statistics only, not the actual data.
Identity & Discovery
Each agent carries two layers of identity. AgentProfile is the lightweight local identity (nickname, role, capabilities) used by in-process agents like the research protocol's Writer and Reviewer. AgentIdentity extends it with hardware capabilities, knowledge statistics, and inference info for network coordination.
What's Broadcast
- Node ID — persistent across restarts (saved to disk), stable address for peer tracking
- Available tools & skills — what this agent can do (used for delegation routing)
- Knowledge statistics — episodic memory count, causal link count, top tools by success rate
- Inference models — what LLMs are loaded, whether inference is available
- Embodiment summary — body modalities and affordances (when embodied)
The identity shares statistics about knowledge, not the knowledge itself. A peer sees "500 episodes, high-confidence grasp patterns" — not the actual memories.
Transport Layer
The PeerChannel implements the same Channel interface as SMS and voice, so mesh messages flow through CommunicationGateway. Sends are queued to a background thread — callers never block on network I/O.
The wire protocol uses MeshMessage — a typed envelope with 24 message types covering discovery, delegation, knowledge sharing, prediction, and inference routing. A protocol_version field ensures peers running different Maxim versions can detect incompatibilities.
Admission Control
MeshAdmissionControl protects agents from flooding. Every incoming message is checked before dispatch. Peers that exceed rate limits accumulate violations with escalating gate durations.
| Trust Level | Rate Limit | How Established |
|---|---|---|
| Verified | 120 msg/min | Pre-shared key |
| Discovered | 60 msg/min | mDNS (future) |
| Remote | 60 msg/min | Tunnel auth |
| Unknown | 20 msg/min | No auth |
Burst detection triggers at 20 messages in 5 seconds. Gate durations escalate: 30s → 2min → 10min → 1hr. Gated peers receive 429 responses and are skipped by the distributed planner.
Knowledge Sharing
Knowledge sharing uses a provider/receiver protocol. Any subsystem can participate by implementing two simple interfaces. The ExperienceBroker routes between them without knowing subsystem specifics.
Built-in Adapters
Causal Links
Wraps NAc. Exports high-confidence links (≥0.6, 3+ observations). Imports with trust-level transfer discount (0.5× for verified, 0.1× for unknown).
Reflections
Wraps Hippocampus. Exports memories with reflection text. Imports with trust-level salience caps (0.5 for verified, 0.15 for unknown).
Motor Programs
Wraps ProgramRegistry. Exports battle-tested programs (3+ executions, ≥50% success). Requires verified/discovered trust. Validates entity paths against local embodiment.
Future adapters (DN thresholds, forward models, contacts) register the same way — no broker changes needed.
Task Delegation
One agent can propose a goal to another. The TaskDelegator picks the best peer by tool availability and success rate. The TaskReceiver evaluates proposals using local NAc predictions and queue depth.
Safety Mechanisms
- Loop detection — Goals carry
delegation_depth(max 2) anddelegation_chain(cycle detection) - Queue limits — Max 5 concurrent delegations per agent
- NAc veto — Rejects if local NAc predicts failure with >70% confidence
- Gated peer skip — Misbehaving peers are excluded from delegation routing
Distributed Planning
The AdaptivePlanner's _tag_delegatable_subgoals() runs after LLM decomposition. It checks each sub-goal against peer capabilities and tags actions with _preferred_node when a peer is better suited:
- Remote-only tools — peer has the tool, we don't → always delegate
- Better success rate — peer has >80% success and local NAc predicts <50% → delegate
Clock Synchronization
The PeerClockEstimator learns per-peer clock offsets using heartbeat round-trip times (NTP-lite). Offsets are smoothed with an EMA and stabilize after ~10 sync points.
The SCN's register_external() corrects incoming TemporalSignatures before registration, ensuring temporal bins align across agents with different system clocks. The phase-normalized indexing is naturally resistant to small offsets — skew only matters when it shifts a memory into the wrong hour bin (>30 minutes).
API Endpoints
The mesh mounts under the existing LeaderProxy, inheriting authentication, rate limiting, and GPU metrics. Three endpoints handle all mesh traffic:
| Endpoint | Method | Purpose |
|---|---|---|
| /v1/mesh/message | POST | Send a MeshMessage to this agent (checked by MeshAdmissionControl) |
| /v1/mesh/peers | GET | List all known peers from PeerRegistry (read-only) |
| /v1/mesh/status | GET | Admission control state — per-peer trust, violations, gate status |
The PeerRegistry bootstraps from the existing peer config file (tunnel URL + API key) or can be populated via mDNS discovery (future). Peers are tracked by persistent node_id, updated on each heartbeat exchange.
Trust Model
| Trust | Knowledge Discount | Motor Programs | Rate Limit |
|---|---|---|---|
| Verified | 0.5× | Accepted | 120/min |
| Discovered | 0.3× | Accepted | 60/min |
| Remote | 0.3× | Rejected | 60/min |
| Unknown | 0.1× | Rejected | 20/min |
Locally-learned knowledge always dominates. Imported data is a prior, not a replacement — as the local agent observes the same patterns, confidence grows organically through its own experience.