Agent Creator Manager

Provision new GPIA agent workspaces and register them for auditing. Each provisioned agent receives:

• A unique identifier and workspace directory
• Scoped helper functions for safe file actions
• A runner template that executes a step plan and writes a report

Quick start

Use the skill with capability=provision and a JSON request:

{
  "agent_name": "LogParser-01",
  "primary_goal": "Clean and format all CSVs in data/export",
  "model_id": "qwen3:latest",
  "skill_categories": ["FileSystem", "Formatting"],
  "ephemeral_mode": false,
  "max_steps": 5,
  "custom_helpers": [
    {"name": "parse_csv", "description": "Parse CSV rows safely"}
  ],
  "output_path": "C:/path/to/output",
  "requester_id": "user",
  "requester_type": "user",
  "parent_agent_id": null
}

You can also call the provisioning tool directly:

python scripts/provision_agent.py --input request.json

Capabilities:

• provision (default): Create a new agent workspace
• list_registry: Return registry entries
• register_runtime: Register an in-memory agent
• provision_skill: Create a skill stub (name + description)

Inputs

Required fields:

• agent_name
• primary_goal
• model_id
• skill_categories (array)

Optional fields:

• ephemeral_mode (bool, default false)
• max_steps (int, default 5)
• custom_helpers (list of helper specs)
• output_path (string)
• requester_id (string)
• requester_type (string: user, model, agent, system)
• parent_agent_id (string or null)
• policy_scope (string, default manual)
• approved / approval_note (used when policy requires approval)
• agent_template (string, default standard, use living_messenger for a persistent loop)
• keep_alive (bool, default false)
• poll_interval (seconds, default 2.0)
• heartbeat_interval (seconds, default 60.0)

Guardrails are defined in config/agent_creator_guardrails.json.

Outputs

Provisioning returns:

• agent_id
• workspace
• output_path
• plan_path
• runner_path
• registry_path