Agent Coordination

The Coordination Challenge

Most trading automation tools operate as isolated systems — a single bot running a single strategy. PerpDesk's 8 agents face a fundamentally different challenge: how do multiple specialized agents coordinate their actions to produce outcomes better than any single agent could achieve alone?

The answer is the Coordination Pipeline — a structured workflow that ensures information flows efficiently between agents and decisions are made with full context.

Pipeline Architecture

The coordination pipeline operates in five phases:

Phase 1: Sense

The sensing phase gathers and processes market data from multiple perspectives:

• Vanta continuously classifies the market regime across all monitored assets
• Luma aggregates and scores signals from technical, on-chain, and sentiment sources
• Meridian monitors macro conditions and updates the systemic risk score

All three sensing agents operate in parallel and share their outputs with the entire team.

Phase 2: Plan

The planning phase translates market intelligence into actionable strategies:

• Riven receives regime data, signal scores, and macro context
• It constructs strategies that align with current conditions
• Strategies are validated against Brix's risk parameters before being proposed
• Multi-leg strategies are coordinated across assets and exchanges

Phase 3: Act

The execution phase turns strategies into positions:

• Koda receives approved strategies from Riven
• It evaluates execution conditions across exchanges
• Orders are placed with optimal timing, routing, and sizing
• Execution reports are generated for every trade

Phase 4: Guard

The guarding phase ensures safety at every step:

• Brix monitors all pipeline activity in real-time
• It validates every proposed action against risk limits
• It can veto strategies, block orders, or force position reductions
• Risk metrics are recalculated continuously

Phase 5: Learn

The learning phase captures insights for continuous improvement:

• Lucid analyzes the outcomes of every decision
• Performance is attributed across agents, strategies, and regimes
• Insights feed back into future sensing, planning, and execution
• Mentor translates learnings into coaching for the trader

Inter-Agent Communication

Agents communicate through a shared context layer that provides:

• State broadcast — Every agent publishes its current state and outputs
• Event bus — Agents can subscribe to specific events from other agents
• Priority messaging — Urgent communications (like Brix risk alerts) take precedence
• Conflict resolution — When agents disagree, structured protocols determine the outcome

Pipeline Timing

The pipeline runs on multiple timeframes simultaneously:

• Real-time — Brix risk monitoring and Koda execution operate continuously
• Fast cycle (seconds) — Vanta regime checks and Luma signal updates
• Medium cycle (minutes) — Riven strategy evaluation and adjustment
• Slow cycle (hours/days) — Lucid performance analysis and Meridian macro updates

This multi-timeframe approach ensures that urgent actions happen immediately while longer-term analysis has time to develop nuanced insights.

Failure Handling

The pipeline is designed for graceful degradation:

• If an agent encounters an error, it publishes a degraded state and the pipeline adapts
• If Vanta can't classify the regime, Riven defaults to conservative strategies
• If Koda can't execute on the preferred exchange, it falls back to alternatives
• If any agent is offline, Brix automatically tightens risk limits as a precaution