Autonomous Mode

Enable autonomous iteration for complex tasks.

What is Autonomous Mode?

Autonomous mode allows Soothe to work independently on complex tasks that require iterative refinement. Instead of stopping after one response, Soothe will:

  1. Create a plan for the goal
  2. Execute the plan step-by-step
  3. Reflect on results after each iteration
  4. Adjust the plan if needed
  5. Continue until the goal is achieved or iteration limit is reached

When to Use Autonomous Mode

Use autonomous mode for tasks that:

  • Require iterative refinement based on results
  • Involve multi-phase research where findings inform next steps
  • Need long-running workflows without manual intervention
  • Decompose into sub-goals that emerge during execution

Examples:

  • “Optimize the simulation parameters across the search space”
  • “Research and document all API endpoints in this codebase”
  • “Build a comprehensive test suite with maximum coverage”
  • “Analyze and improve the performance of this application”

How to Enable

TUI Command 

# In TUI
/autopilot Optimize the simulation parameters

# With custom iteration limit
/autopilot 15 Research and improve model performance

Autopilot Command 

# Dedicated autopilot command
soothe autopilot run "Research quantum error correction advances"

# The --max-iterations flag is accepted but ignored.
# Configure the iteration limit via daemon config: agent.autopilot.max_iterations
soothe autopilot run "Build a web scraper"

Progress Events

During autonomous execution, you’ll see progress events:

  • soothe.cognition.strange_loop.started - StrangeLoop execution began
  • soothe.cognition.strange_loop.completed - StrangeLoop execution finished
  • soothe.cognition.goal.created - New goal created
  • soothe.cognition.goal.completed - Goal achieved
  • soothe.cognition.goal.failed - Goal failed

Configuration

Set defaults in your configuration file:

agent:
  autopilot:
    # Enable autopilot scheduling loop (starts on daemon startup for 24/7 operation)
    enabled: false

    # Maximum iterations per autopilot thread (default: 10)
    max_iterations: 10

    # Maximum retries per goal on failure (default: 2)
    max_retries: 2

    # Wall-clock budget per dispatched goal (default: 1209600 = 14 days)
    goal_deadline_seconds: 1209600

Set goal_deadline_seconds: null to disable autopilot deadline enforcement (not recommended for unattended 24/7 runs).

Environment Variables 

export SOOTHE_AGENT_AUTOPILOT_ENABLED=true
export SOOTHE_AGENT_AUTOPILOT_MAX_ITERATIONS=15
export SOOTHE_AGENT_AUTOPILOT_MAX_RETRIES=3
export SOOTHE_AGENT__AUTOPILOT__GOAL_DEADLINE_SECONDS=1209600

How It Works

Iteration Loop

  1. Plan Creation: Soothe creates a plan for your goal
  2. Step Execution: Each plan step is executed
  3. Reflection: Results are analyzed
  4. Plan Adjustment: Plan is updated if needed
  5. Continuation: Loop continues until goal achieved or limit reached

Stopping Conditions

Autonomous mode stops when:

  • The goal is achieved
  • Maximum iterations reached
  • Wall-clock deadline exceeded (goal_deadline_seconds, default 14 days)
  • User cancels with /cancel or Ctrl+C
  • Critical error occurs

Examples

Optimization Task 

soothe autopilot run "Optimize the database queries for better performance"

Soothe will:

  1. Analyze current query performance
  2. Identify bottlenecks
  3. Optimize queries
  4. Measure improvements
  5. Iterate until satisfied

Research Task 

soothe autopilot run "Research and document the best practices for REST API design"

Soothe will:

  1. Research REST API best practices
  2. Gather multiple sources
  3. Synthesize findings
  4. Create documentation
  5. Refine based on gaps

Development Task 

soothe autopilot run "Build a comprehensive test suite for the authentication module"

Soothe will:

  1. Analyze the authentication module
  2. Identify test scenarios
  3. Write tests
  4. Run tests and check coverage
  5. Add more tests iteratively

Monitoring Progress

TUI

In the TUI, you’ll see:

  • Current iteration number
  • Active goals
  • Plan progress
  • Subagent activity
  • Tool usage

Headless Mode

With --format jsonl, each progress event is a JSON object:

{"type": "event", "event_type": "soothe.cognition.strange_loop.started", "data": {"iteration": 1}}
{"type": "event", "event_type": "soothe.cognition.goal.created", "data": {"goal_id": "goal_1", "description": "..." }}

Best Practices

  1. Clear Goals: Provide specific, measurable objectives
  2. Reasonable Limits: Start with default 10 iterations
  3. Monitor Progress: Check in on long-running tasks
  4. Cancel if Needed: Use /cancel or Ctrl+C to stop