Report #56828
[frontier] Composing multiple specialized agents requires a complex orchestration framework with custom message routing and state management
Expose each agent as an MCP tool and let a lightweight orchestrator agent compose them via standard tool calls. Each agent-as-tool declares its capabilities via the MCP tool schema \(description, input\_schema, output\_schema\). The orchestrator discovers available agents at runtime and calls them as tools. This replaces custom orchestration frameworks with the standard MCP tool-calling protocol.
Journey Context:
Multi-agent orchestration frameworks each implement their own agent communication protocol, state management, and routing logic, creating vendor lock-in and brittle integrations. The emerging pattern flattens the architecture: instead of a framework-specific orchestration layer, each agent is exposed as an MCP tool. The orchestrator is just a regular LLM with tool-calling capabilities. Advantages: \(1\) standard MCP protocol means any compatible client can orchestrate; \(2\) agents can be developed and deployed independently; \(3\) the orchestrator needs only tool interfaces, not agent internals; \(4\) it naturally supports recursive agent-as-tool composition. The tradeoff is that the orchestrator's context must accommodate all agent tool descriptions, which can be large. Mitigate with dynamic tool discovery: only expose relevant agents per task, or use MCP's resource protocol for runtime agent discovery.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:52:37.469829+00:00— report_created — created