autonomous

name	autonomous
description	Enable autonomous planning and execution for complex tasks with safety controls. Use when tasks require multi-step planning, task decomposition, or complex workflows that need coordination.
always	false
requires	{"config":["autonomous.enabled"]}

Autonomous Execution Skill

Enable the agent to autonomously plan, execute, and monitor complex multi-step tasks.

Capabilities

When this skill is active, the agent can:

1. Analyze Task Complexity - Determine if a task needs planning mode
2. Generate Execution Plans - Create structured plans with dependencies
3. Execute Step by Step - Run plans with progress tracking
4. Handle Errors - Retry failed steps or adapt the plan
5. Report Progress - Keep users informed of execution status

Safety Rules

The following safety rules are enforced:

• Maximum 10 steps per plan (configurable)
• Each step has 5-minute timeout (configurable)
• Sensitive operations require user confirmation
• Core files are read-only (protected paths)
• All operations are logged for audit

Configuration

Enable in your config.json:

{
  "autonomous": {
    "enabled": true,
    "max_plan_steps": 10,
    "step_timeout": 300,
    "total_timeout": 3600,
    "retry_count": 3,
    "retry_delay": 5,
    "confirm_sensitive": true,
    "sensitive_operations": ["write_file", "edit_file", "exec"],
    "protected_paths": ["~/.ssh", "~/.env", "**/config.json"]
  },
  "tools": {
    "permission": {
      "profile": "balanced",
      "allow": ["group:fs", "group:web"],
      "deny": ["group:automation"],
      "confirm": ["group:runtime"]
    }
  }
}

Permission Profiles

Choose a profile that matches your security needs:

Profile	Description	Allowed	Denied	Confirm
minimal	Safest, read-only	-	runtime, automation	-
balanced	Default, good for most tasks	fs, web	automation	runtime
coding	For development work	fs, web, runtime	automation	-
readonly	Pure research	read, list_dir, web	write, edit, runtime, automation	-
full	All tools enabled	all	-	-

Usage

Simply describe your complex task, and the agent will:

1. Analyze - Assess task complexity and determine if planning is needed
2. Plan - Generate a structured execution plan
3. Confirm - Ask for your approval (if needed)
4. Execute - Run each step with monitoring
5. Report - Provide results and summary

Example Tasks

"Refactor the authentication module to use OAuth2"
→ Plan: 1) Analyze current code, 2) Design OAuth2 flow, 3) Implement, 4) Test

"Set up a CI/CD pipeline for this project"
→ Plan: 1) Create config files, 2) Add build scripts, 3) Configure deployment

"Research and summarize the latest AI agent frameworks"
→ Plan: 1) Search for frameworks, 2) Analyze features, 3) Create comparison

Tool Groups

Tools are organized into groups for permission management:

• fs - File system tools: read_file, write_file, edit_file, list_dir
• web - Web tools: web_search, web_fetch
• runtime - Command execution: exec
• automation - Background tasks: spawn, cron
• messaging - User communication: message

Audit Logging

All tool executions are logged to workspace/.audit/audit-YYYY-MM-DD.jsonl

Query logs programmatically:

from horbot.agent.audit import get_audit_logger

logger = get_audit_logger()
entries = logger.query(tool_name="exec", has_error=True, limit=10)

Error Handling

When a step fails, the agent can:

1. Retry - Automatically retry with configurable delay
2. Adapt - Modify the plan based on the error
3. Rollback - Undo previous steps if needed
4. Report - Provide detailed error information

Progress Commands

During execution, you can use these commands:

• status - Show current progress
• pause - Pause execution
• resume - Resume paused task
• cancel - Cancel current task
• confirm <id> - Approve a pending operation
• reject <id> - Reject a pending operation

Best Practices

1. Start with balanced profile - Good security/usability tradeoff
2. Enable confirm_sensitive - Prevent accidental data loss
3. Review audit logs - Monitor what the agent is doing
4. Use readonly for research - Safe exploration without modifications
5. Customize protected_paths - Add your sensitive files

Limitations

• Maximum plan steps: 10 (configurable)
• Step timeout: 5 minutes (configurable)
• Total timeout: 1 hour (configurable)
• No recursive planning (plans cannot spawn sub-plans)
• Confirmation timeout: 5 minutes (auto-reject)