"Agent Messaging & Orchestration Roadmap (Aspirational)"

Agent Messaging & Orchestration Roadmap (Aspirational)

This document outlines the aspirational goals for the Vox Distributed Execution Intelligence (DEI) orchestrator and agent-to-agent (A2A) messaging architecture, tracking toward state-of-the-art 2026 multi-agent patterns.

1. Context Management Evolution

Current State: Context is primarily bounded by file selections, explicit @mentions, and static chat history keys. Aspirational Goals:

  • Continuous Context Engineering: Move beyond static prompt injection. Introduce automatic real-time context summarization where long-running agent threads compress their episodic memory into semantic checkpoints.
  • Multimodal State Integration: Support the injection of UI visual snapshots and multimodal telemetry natively in ChatMessage constructs, preventing agents from becoming text-blind to DOM or pixel-level changes.
  • Context Routing: Implement policies that automatically "shed" irrelevant history when an agent shifts execution domains (e.g., from database debugging to UI CSS tweaking) -> save token budgets and prevent hallucination bleed.

2. Multi-Agent Topologies & Orchestration

Current State: Tasks are routed to the most capable single agent based on affinity (vox-orchestrator's routing service). Aspirational Goals:

  • Specialized "Agent Pods": Break down monolith tasks into sub-delegations using a hierarchical task network (HTN). Assign specialized agents (Planner, Executor, Verifier, Researcher) -> specific nodes instead of relying on general-purpose code-gen agents.
  • Dynamic Handoff/Triage (Delegation Pattern): An agent can unilaterally pause execution to issue an A2A RPC requesting help from an agent with higher Trust or specific tool permissions (e.g., a "Security Agent" for signing commits or handling API tokens).
  • Parallel Analysis (Map-Reduce): The Orchestrator should support spawning N ephemeral agents to analyze independent files concurrently across the mens, gathering the results via an accumulator agent.

3. Advanced Memory & Socrates Integration

Current State: vox_chat_message and vox_memory_search share a unified retrieval trigger that prefers hybrid BM25 + vector search and falls back deterministically when embeddings/DB are unavailable. Broader autonomous contradiction-resolution orchestration remains aspirational. Aspirational Goals:

  • Autonomous Subconscious Recall: All LLM entrypoints should automatically run a low-latency vector-BM25 hybrid query against the Codex memory block using the user's prompt as the latent space seed. High-confidence facts (score > 0.85) should silently append to the preamble, fulfilling the "agent knows when to look" imperative.
  • Contradiction Resolution Agents: If the MemorySearchEngine detects a potential_contradiction, the Orchestrator should automatically pause the fast-path pipeline and insert a "Resolution Re-plan" task, spawning an investigative agent to resolve the factual split before the primary agent generates code.

4. System Governance as an 'OS' Layer

Current State: Orchestrator enforces basic limits (max_agents, stale_threshold_ms, lock contention). Aspirational Goals:

  • Structured Orchestration Transitions: Formalize task execution into a state machine: Understand -> Plan -> Act -> Evaluate. Currently, agents can loop infinitely unless gated. This OS-level transition forces an episodic commit at each boundary.
  • Standardized A2A Protocol Alignment: Expose the internal MessageBus to conform fully with emerging 2026 standards like Google's Agent-to-Agent (A2A) protocol or Anthropic's Model Context Protocol (MCP) multi-agent routing extensions, allowing Vox mens nodes to interoperate with non-Vox, third-party agents running on external infrastructure.

Next Steps for Build-out

  1. Implement basic session-isolated history in vox-mcp (Immediate).
  2. Extend chat retrieval into task-level replan orchestration when contradiction hints are detected (Immediate).
  3. Draft the HTN topology spec for vox-orchestrator/src/queue.rs (Q3 2026).
  4. Build the PodManager to enforce specialized agent teaming (Q4 2026).