Types of Agents

Too little autonomy (approve every click) makes agents useless. Too much autonomy (unsupervised financial transactions) is dangerous. Production agents need calibrated autonomy levels that match task risk, user trust, and regulatory requirements.

Autonomy levels are like a driver's license progression — learner's permit (supervised), independent driver (most tasks), commercial license (high-stakes operations with extra checks).

Scenario

Email agent: reading inbox = autonomous. Drafting reply = confirm (show draft, user approves). Sending to external contacts = supervised (manager reviews). Deleting emails = blocked (not in allowlist).

Solution

In Agent Foundations, apply Types of Agents to this scenario: Email agent: reading inbox = autonomous. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Types of Agents (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Types of Agents happens in the code.

1AUTONOMY_POLICIES = {2    "read_email": "autonomous",3    "draft_email": "confirm",4    "send_email": "supervised",5    "delete_email": "blocked",6    "query_database": "autonomous",7    "update_database": "confirm",8    "transfer_funds": "blocked",9}10
11def execute_with_autonomy(tool_name: str, args: dict, user_id: str) -> dict:  # define a reusable function12    policy = AUTONOMY_POLICIES.get(tool_name, "confirm")13
14    if policy == "blocked":15        return {"status": "blocked", "reason": f"{tool_name} is not permitted"}  # return the result16
17    if policy == "confirm":18        approval = request_human_approval(user_id, tool_name, args)19        if not approval:20            return {"status": "rejected"}  # return the result21
22    result = execute_tool(tool_name, args)23    audit_log(user_id, tool_name, args, result)24    return {"status": "success", "result": result}  # return the result

• Full autonomy on all tools — dangerous in production
• Requiring approval for every action — agent is useless
• No audit trail — can't investigate incidents
• Static autonomy levels — not adapting to agent track record
• Ignoring regulatory requirements for human oversight

Quick recap — the most important points from this module.