Страницы · Claude Agents Team

/meet — streamed multi-agent meeting

Gather every agent on the active team around a topic. Each speaks from their domain — constraint check, top 3 risks, dependencies, open questions, effort. The orchestrator consolidates to a global top-10, batches open questions for the client, and writes minutes to disk.

The skill is documented in plugins/agents-team/skills/meet/SKILL.md.

When to run it

Before approving a major plan in .plans/.
After a significant client directive that reshapes the architecture.
Before committing to a new service, tool, or vendor.
When the orchestrator needs cross-domain consensus on a risk or trade-off.

Quick path

bash

What you get

A single Markdown file at .meetings/YYYY-MM-DD-{topic-slug}.md. If a meeting already exists for the same topic + date, the skill appends -2, -3, … rather than overwriting.

markdown

How it works

Resolve the topic — if it's a path under .plans/ or .requests/, load it.
Discover the team — Glob .claude/agents/*.md. Skip disable-model-invocation: true and background: true agents (background monitors don't speak in meetings).
Dispatch in parallel — one Agent call per attendee, each with a fixed structured prompt.
Stream — each agent's response lands as it arrives; do not buffer.
Consolidate — global top-10 risks, disagreements, open questions, recommendation.
Write the minutes file. Print the path + recommendation as a one-liner.

Cost expectations

Wall-clock: 1–3 minutes for a 5-agent team; 5–10 minutes for a 15-agent team. Parallel dispatch keeps this bounded.
Tokens: ~200 input + ~200 output per attendee, plus the orchestrator on top.

Anti-patterns

Treating one agent's view as the team's view — the whole point of /meet is the cross-domain check.
Buffering all responses then dumping at the end — defeats streaming.
Letting agents respond out-of-format — ask them to redo (one retry, then move on).
Overwriting prior minutes — same topic + date appends -N, never overwrites.