Tutti — Multi-Agent Orchestration

Orchestrate a team of AI coding agents from a declarative tutti.toml config. Launch agents in isolated git worktrees, run verification workflows, track token usage, and manage context handoffs — all through a single CLI.

When to use this skill

Use when the user asks you to:

• Launch, monitor, or stop a team of AI coding agents
• Run or verify automated workflows across agents
• Dispatch prompts to agents with auto-start and output capture
• Land agent work back to the main branch or open PRs
• Check agent status, health, or capacity usage
• Generate or apply context handoff packets
• Coordinate multi-agent development workflows

Prerequisites

1. tt binary installed and on PATH (install from https://github.com/nutthouse/tutti)
2. tmux installed
3. python3 available
4. A tutti.toml config file in the workspace root

Always run preflight checks before starting a workflow:

python3 tutti_openclaw.py doctor_check

Actions

All actions go through the wrapper script. Every action returns a consistent JSON envelope:

{
  "ok": true,
  "action": "action_name",
  "command": ["tt", "..."],
  "exit_code": 0,
  "data": {},
  "stdout": "",
  "stderr": ""
}

Lifecycle

Action	Command	Purpose
doctor_check	python3 tutti_openclaw.py doctor_check	Preflight: verify tools, config, and environment
launch_team	python3 tutti_openclaw.py launch_team	Launch all agents defined in tutti.toml
launch_agent	python3 tutti_openclaw.py launch_agent <name>	Launch a single agent
send_prompt	python3 tutti_openclaw.py send_prompt <agent> <prompt...> [--auto-up] [--wait] [--output]	Send a prompt to an agent with optional auto-start, wait-for-idle, and output capture
team_status	python3 tutti_openclaw.py team_status	Read agent states from .tutti/state/
agent_output	python3 tutti_openclaw.py agent_output <name> --lines 50	Peek at an agent's terminal output
stop_agent	python3 tutti_openclaw.py stop_agent <name>	Stop a single agent
stop_team	python3 tutti_openclaw.py stop_team	Stop all agents

Workflows

Action	Command	Purpose
list_workflows	python3 tutti_openclaw.py list_workflows	Discover available workflows
plan_workflow	python3 tutti_openclaw.py plan_workflow <name> [--strict]	Dry-run a workflow
run_workflow	python3 tutti_openclaw.py run_workflow <name> [--agent <a>] [--strict]	Execute a workflow
verify_team	python3 tutti_openclaw.py verify_team [--workflow <w>] [--strict]	Run verification workflow
read_verify_status	python3 tutti_openclaw.py read_verify_status	Read last verification result

Git Operations

Action	Command	Purpose
land_agent	python3 tutti_openclaw.py land_agent <agent> [--pr] [--force]	Land an agent's branch back to current branch, or open a PR

Handoffs

Action	Command	Purpose
generate_handoff	python3 tutti_openclaw.py generate_handoff <agent> [--reason <r>]	Capture agent context to a packet
apply_handoff	python3 tutti_openclaw.py apply_handoff <agent> [--packet <path>]	Inject a handoff packet into an agent
list_handoffs	python3 tutti_openclaw.py list_handoffs [--agent <a>] [--limit 20]	List available handoff packets

Permissions

Action	Command	Purpose
permissions_check	python3 tutti_openclaw.py permissions_check <cmd...>	Check if a command is allowed by policy

Workflow step types

Workflows in tutti.toml support these step types:

Type	Purpose	Key fields
prompt	Send text to an agent session	agent, text, inject_files, wait_for_idle, wait_timeout_secs
command	Execute a shell command	run, cwd, timeout_secs, fail_mode
ensure_running	Start an agent if not already running	agent, fail_mode
workflow	Execute another workflow as a nested step	workflow, agent, strict, fail_mode
land	Land an agent's branch	agent, pr, force, fail_mode
review	Send an agent's diff to a reviewer	agent, reviewer, fail_mode

Prompt steps support inject_files — an array of workspace-relative file paths that are copied into the agent's worktree before the prompt is sent. This enables stateful context passing between agents (e.g., injecting a snapshot JSON produced by another agent).

Nested workflow steps enable composition: observe → dispatch → fix → verify → land as a chain of workflow invocations.

Execution pattern

Follow this sequence for orchestrating a workspace:

1. Preflight — doctor_check. Stop and report if non-zero.
2. Launch — launch_team or launch_agent <name>.
3. Monitor — team_status and agent_output <name> to observe progress.
4. Dispatch — send_prompt <agent> "do something" --auto-up --wait --output to dispatch work and capture results.
5. Workflow — list_workflows to discover, then run_workflow <name>.
6. Verify — verify_team --strict for gate-style quality checks.
7. Land — land_agent <agent> to cherry-pick work, or land_agent <agent> --pr to open a PR.
8. Handoff — generate_handoff <agent> when context is high, apply_handoff <agent> to resume.
9. Stop — stop_team or stop_agent <name> when done.

Failure handling

• Non-zero exit: Surface the action, command, and stderr from the JSON envelope. Do not retry blindly.
• Verify warnings (non-strict): Report as warning. Include data from read_verify_status.
• Missing state files: Treat as transient — retry up to 3 times with short delays. If still missing, the workspace may not have been launched.
• Auth failures: If stderr contains auth errors, stop and escalate to the user. Do not retry auth failures.
• Agent not running: Use --auto-up on send_prompt to automatically start agents on demand rather than failing.

Configuration override

If tt is not on PATH or you need a specific version:

python3 tutti_openclaw.py --tt-bin /path/to/tt doctor_check
# or via environment variable
TUTTI_BIN=/path/to/tt python3 tutti_openclaw.py doctor_check

Rules

• Always run doctor_check before any launch or workflow operation.
• Never retry auth failures — escalate to the user immediately.
• Prefer team_status (reads state files directly) over agent_output for status checks.
• Use --strict flag on verify_team and run_workflow when results gate further actions.
• Use --auto-up on send_prompt when the target agent may not be running.
• Use --output on send_prompt to capture the agent's response for programmatic verification.
• Use --json output from tt commands when you need structured data (the wrapper handles this automatically).
• Do not parse stdout text output — always use the data field from the JSON envelope.