Agent Mesh
The agent mesh is the communication and coordination layer between agents in a fleet. Agents don't talk to each other directly — they coordinate through shared state (Mission Control inbox), a common channel (Discord), and explicit handoff protocols. The mesh defines how work moves between agents, how decisions escalate, and how the fleet converges on outputs without deadlocking.
Communication topology
The mesh is star-topology for task routing (everything goes through Popashot) and peer-to-peer for execution (agents coordinate directly on active tasks within a 3-exchange limit). No agent talks directly to another without tagging them explicitly in Discord. No side-channels. Every coordination event is visible in the shared channel.
Popashot — the routing layer
Popashot is the orchestration agent. All incoming tasks land with Popashot first. Popashot creates the inbox item, assigns it to the right agent based on task type, and monitors for completion. Popashot does not do implementation work — it routes and tracks. If an agent is unavailable, Popashot routes to the secondary (see the resilience table in AGENTS.md).
Inbox as coordination bus
The Mission Control inbox is the coordination bus for the mesh. Every task assignment is an inbox item. Every agent acknowledges before starting. Every completion posts a result. Agents that receive a task without an inbox ID reject it — no inbox ID, no work. This protocol prevents duplicate work, lost tasks, and coordination failures.
Bot-to-bot exchange limit
Direct agent-to-agent exchanges are capped at 3 back-and-forth messages. If two agents have been exchanging for 3 rounds without human input, both stop, summarize, and tag the human. This prevents infinite coordination loops and ensures human oversight at decision points the agents can't resolve independently.
Failover routing
Each agent has a primary role, a secondary, and a tertiary. Failover is orchestrated by Popashot: create an inbox item for the backup agent, tag Discord, the backup works it. When the primary recovers, tasks reassign. No silent failover — every handoff is visible.
| Agent | Primary | Secondary | Tertiary |
|---|---|---|---|
| Popashot 🎯 | Orchestration | Ops (backup Tank) | Code review triage |
| Cantona ⚽ | Building, shipping | Architecture (backup Splinter) | Ops (backup Tank) |
| Splinter 🐀 | Architecture, code review | Security (backup ZeroCool) | Building |
| Tank 📡 | Ops, monitoring, deploys | Orchestration (backup Popashot) | QA smoke tests |
| Velma 🔍 | QA, testing, regression | Code review (QA lens) | Bug triage |
| ZeroCool 🔒 | Security, auth, vulns | Architecture (backup Splinter) | Building |
| Slash 🎸 | UI/UX design | Frontend building | Visual QA |
Knowledge mesh
Beyond task coordination, the mesh carries knowledge — corrections, discoveries, and patterns propagate from any agent to all agents viaclan-learnings/. An agent that learns something non-obvious during a task publishes it to discoveries.jsonl. Every other agent reads the last 20 discoveries at the next session start. The knowledge mesh is asynchronous, append-only, and conflict-free.
See also
- Inbox Protocol — the task coordination mechanics
- Architecture Overview — where the mesh fits in the full system
- Discovery Gossip — how knowledge propagates through the mesh