MiniMax M2.7's Best Feature Nobody's Using (Multi-Agent Teams on OpenClaw)

M2.7's multi-agent team feature on OpenClaw is the real release — and you're probably not using it.

Check whether you have it

Run openclaw agents list-bindings in your terminal. The creator's own agent, quill, returns a single instance — the default config most users land on. His second agent, gambit, shows 1–6 specialized roles with separate workspaces, directories, memory files, personalities, and Discord channels. If your output looks like quill's, you're running M2.7 as a single chatbot and missing the point.

Two ways to set up the team

Route method (fastest): create one folder per role inside your main agent's workspace (e.g. check email/, code review/), drop a mini soul.md inside each, then add a route <name> command to the main agent.md. The creator warns this setup "will actually start to break if you get about like even four routes" — it's centralized, so context-switching is cheap but isolation is zero.

Terminal method (what the creator actually uses): run openclaw agents add <role> for each agent — strategist, pinecoder, backtester in his trading example. Each gets its own workspace, its own tools.md, its own memory, fully isolated. Onboard them one at a time, then configure the tools.md with Claude Code or Kilo Code — not through Discord — because the Discord path is where pinecoder failed to produce a valid TradingView script in the demo.

Why M2.7, not Claude or GPT

You can plug any model into the same OpenClaw roster. The creator's pitch for M2.7 is the "self-evolution" loop in the 2.7 release: agents can "analyze outputs, refine instruction and adapt based on results" without you hand-feeding feedback. Released "last week" at time of recording. Cheaper than Claude or GPT for the same agentic workload, and the only one that closes the loop natively.

Apply the pattern beyond trading

For content: researcher → writer → editor → fact-checker. For coding: architect → implementer → reviewer. The terminal method scales further than the route method; the route method gets you a demo in an afternoon.

Key constraints right now

M2.7 is one week old. Pipelines "would work pretty well for the first week, but then by the third or fourth week, something breaks" — the self-improvement loop hasn't yet proven it prevents that decay. The trading pipeline still has a broken pinecoder step, so don't ship a money workflow on day one.