← All work
evolving · 2 min read

Agentic MCP Teaming

Multi-agent coordination framework where named specialists collaborate via a shared MCP bus: parallel review, unanimous consensus on design, and isolated Git worktree implementation.

TypeScriptMCPGenAIAgents
Status
evolving
Active since
Mar 2026
Last updated
June 2026
Changes shipped
13

The orchestration layer behind multi-agent reviews and builds; the engine Steward drives.

Agentic MCP Teaming turns a single assistant into a coordinated team. Named specialists, each with a focused role, collaborate over a shared Model Context Protocol bus, reviewing work in parallel and converging on a design only once they agree, rather than one model guessing alone.

How it works

Agentic MCP Teaming coordination flow Named agents connect to a single MCP coordinator that drives a phased workflow. Every round, all reviewers run in parallel against the same snapshot, then the coordinator evaluates consensus. Approved tasks are implemented in isolated Git worktrees. Participants Workflow Build Agents specialists MCP bus one coordinator Review in parallel Consensus unanimous Worktrees isolated

Everything is one MCP server: tool access, agent invocation, workflow control, and consensus decisions all flow through it, so every action lands in one audit trail and any MCP-capable client can join. The workflow advances through phases (proposal, design, spec, task, implementation) and at each boundary all reviewers run in parallel against a single captured snapshot before any feedback is shared. Approved tasks are implemented in their own Git worktree: a patch is generated, applied, reviewed, and only then merged back.

Key decisions

  • One MCP bus for everything: there are no side channels; every tool call, agent trigger, and decision is a tool call on the same server, which is what makes the audit trail complete and adding a participant a matter of registration, not integration.
  • Independent first-pass review: all reviewers see the same snapshot and respond before any feedback is exchanged, so no agent anchors another’s judgement; only then is consensus evaluated.
  • Consensus and human-approved are distinct outcomes: the system records whether agents truly agreed or a human broke a deadlock, keeping the audit trail honest about how each artifact was resolved.
  • Isolation by filesystem, not locks: each task implements in its own worktree, so parallel work can’t collide on a shared checkout; conflicts surface as a rebase-and-re-review, not a race.

Recent changes

  • Coordinator streamable http
  • Vscode coordinator ui
  • Opsx review gate
  • Coordinator tls transport
  • Coordinator serve mode
  • Cmux backend integration