Orchestration

How lead agents coordinate their team using the orchestrator MCP tools.

The orchestrator

When a lead agent has reporters, muxAI automatically injects two MCP tools and a team roster into the lead’s system prompt. The lead doesn’t need to know internal tool names. It just sees instructions like:

You have direct reports. Use the orchestrator tools to coordinate them.

Tools

run_team

Invokes all reporters in parallel. Each reporter runs their own configured default prompt. Returns every reporter’s name, role, status, and full output. Use this to gather initial findings from the whole team.

ask_reporter

Invokes one reporter by name. Optionally pass a task to override their default prompt. Use this for targeted follow-ups, re-runs, or when you need to provide specific context (e.g. a chart URL or asset pair).

Decision recall

Agents with reviewDecisions enabled (trading templates by default) also get get_my_decisions. It returns the agent’s recent structured decisions along with any user-marked outcomes (win / loss / custom label, plus free-form fields like pnl or note) and a running win/loss tally. The injected system prompt nudges the agent to reflect on past outcomes before its next decision.

For updating the conviction on a specific prior decision rather than starting a fresh one, see Re-examination — a separate user-triggered flow that anchors a new full-council run to a parent decision and produces a re-examination card with an updated conviction score.

When to use which

Scenario	Tool	Why
Initial data gathering	run_team	Fires all reporters in parallel, one turn
Follow-up question to one analyst	ask_reporter	Targeted, with optional task override
Lead provides per-reporter context	ask_reporter	Each reporter gets tailored input (asset, chart, timeframe)
Resolve conflicting findings	ask_reporter	Re-run the analyst whose data you want to verify

Name matching

ask_reporter requires the exact agent name (not the role). If the lead gets the name wrong, the orchestrator returns an error listing available reporter names. The lead self-corrects on the next turn.

Reporter "technical-analyst" not found.
Available reporters: Technical Analyst, News Analyst, Data Analyst

This is by design. The error message is informative enough for the agent to recover automatically.

Turn cost

Each tool call and response consumes one turn. Understanding turn cost helps you set appropriate maxTurnsPerRun values:

Action	Turns used
run_team (success)	1
ask_reporter x3 (all succeed)	3
ask_reporter with wrong name + retry	2
Wrong names for 3 reporters + retry all	4-6

If the lead calls ask_reporter individually for each reporter with wrong names, that’s one wasted turn per reporter before the retry. With 3 reporters, that’s 3 wasted turns. The lead template ships with maxTurnsPerRun: 30 to absorb this overhead.

Auto-injected context

When an agent has reporters, muxAI appends a block to the system prompt:

## Your Team

You have direct reports. Use the orchestrator tools to coordinate them.

- Invoke all reporters in parallel: call `mcp__orchestrator__run_team`
- Invoke one reporter by name: call `mcp__orchestrator__ask_reporter`

Your reporters:
- News Analyst (news-analyst) - Assess sentiment and impact of crypto news
- Technical Analyst (technical-analyst) - Identify directional bias from charts

This means your SKILL.md doesn’t need to list orchestrator tools. Just write natural language about what the lead should do. The tool details are injected automatically.