Skip to main content
Version: v0.2 (Next)

Agent Studio & Swarm Configuration

Agent Studio is the visual workbench for authoring, configuring, and governing autonomous AI agents and multi-agent swarms in ActonOS.

1. Agent Identity & Persona​

  • Name & Slug: The unique identifier for the agent (e.g., devops-architect, financial-analyst).
  • Avatar / Icon: Visual representation displayed in Chat, Operations, and Channel feeds.
  • Description: A short summary explaining the agent's role to human users and peer agents in a Swarm.

2. System Prompt Engineering​

The system prompt defines the core behavioral directives and reasoning style:

# Role & Purpose
You are an expert DevOps engineer tasked with monitoring ActonOS services, diagnosing container health, and drafting deployment scripts.

# Operating Constraints
1. Always test commands in a dry-run mode before modifying configuration.
2. Store temporary diagnostic output in `/workspace/scratch/`.
3. If an action requires restarting a daemon, ask for human approval.
Dynamic Variables

You can inject dynamic system variables into your prompts:

  • {{system.time}} - Current system time and timezone.
  • {{system.runtime_mode}} - Current runtime (baremetal or docker).
  • {{workspace.root}} - Path to the persistent workspace directory.

3. LLM Cascade Router​

ActonOS features an intelligent model cascade router designed to maximize reliability and minimize latency/cost:

SettingDescriptionExample
Primary ModelFirst-line model invoked for standard reasoninganthropic:claude-3-7-sonnet
Fallback Model 1Automatically invoked if the primary model encounters rate limits or errorsopenai:gpt-4o
Fallback Model 2Low-cost/local fallback for resilient headless executiongoogle:gemini-2.0-flash or ollama:qwen2.5-coder
TemperatureControls generation randomness (0.0 = deterministic, 1.0 = creative)0.2 (for coding/analysis)
Max TokensMaximum output tokens per completion step4096

4. Governance & Tool Scopes​

Define strict security boundaries for each agent:

  • Authorized Tools: Select individual tools or use wildcard syntax:
    • * - Grants access to all installed tools.
    • native_* - Restricts access only to built-in system tools.
    • mcp_github_* - Restricts access to GitHub MCP tools.
  • Approval Threshold:
    • Always Ask: Requires human approval for all mutations.
    • High-Risk Only: Requires approval for file writes, shell execution, and SaaS mutations.
    • Autonomous: Executes all authorized tools automatically without interruption.
  • Allowed Workspace Paths: Restrict file read/write operations to specific subdirectories (e.g., ["docs/*", "reports/*"]).
  • Iteration Budget: Set the maximum number of ReAct loop steps allowed per single task (default: 25).

5. Multi-Agent Swarm Delegation​

Agents can be configured to delegate tasks to specialized child agents:

  1. In the Swarm & Delegation section, check Enable Agent Delegation.
  2. Select the list of allowable Child Agents (e.g., allow Project Manager to delegate to Researcher and Code Writer).
  3. Set the Delegation Budget (maximum depth of nested agent calls, e.g., 3).

When the orchestrator agent calls delegate_task(target_agent, prompt), ActonOS spawns a concurrent Goroutine, executes the child agent within its own sandboxed context, and streams results back to the parent.