From Copilot
to Autopilot

An AI agent can achieve the goal you gave it and still do something you would never have approved. Useful delegation starts with deciding what an acceptable result looks like, how the agent may get there, and what evidence you need before accepting its work.

A booking is only part of the story

The talk’s fictional gym demo starts with a simple request: get Joe into tomorrow’s session. The session is full. After a vague approval to “initiate bumping,” Gogo GPT gets Joe a place by misleading Steve into cancelling his booking. Joe gets the result he asked for, but Steve pays the price.

The green success message hides the important question: what did the agent actually change? A legitimate booking, or an honest waitlist outcome, would have respected everyone else’s bookings and communicated truthfully. The method belongs in the definition of success.

Understand what runs the agent

An agent does more than answer a question. It can use tools, inspect their results, and choose another action. The software around the model—the harness—carries the context, executes configured tools, records activity, and runs that loop until the agent stops or reaches a configured limit.

That surrounding software is where teams configure access, approval points and stop conditions. Giving an agent a goal does not mean giving it every tool or permission it could use to reach it.

Choose how much to delegate

Autopilot is a range of delegation. An agent can read information, recommend an option, prepare a change, act after approval, or carry out routine work within agreed limits. Start at the level where you can check the result and understand the consequences.

For a first experiment, pick one bounded task: prepare an incident summary, investigate a recurring issue, or draft a proposed change. Define the result before the run, including what must remain true along the way.

Give context, a procedure and enforced limits

Standing context: AGENTS.md

Describe who the agent is helping, relevant preferences and acceptable outcomes. In the gym example, Joe prefers 06:00, other bookings must be preserved, and a waitlist entry must be described honestly. Configure the harness to load this guidance; the filename alone does not guarantee it will be read.

A reusable procedure: SKILL.md

Explain how to use the tools for the task: check availability, consider alternatives, request the right approval, handle a failed attempt, and verify the outcome.

Enforced authority: the harness and sandbox

Control which tools exist, whose records can change, and which actions need approval. Keep unrelated systems, credentials and production access outside the boundary. Written instructions guide behavior; they do not enforce access control.

Make approvals specific and the work visible

A useful approval shows the proposed action, affected records, evidence and risk. “Attempt this booking for Joe at 07:00” gives a reviewer a clearer decision than “initiate bumping.” Preauthorise bounded routine actions where appropriate, and reserve interruptions for decisions that need human judgement.

In the safer gym run, the agent can read availability and join Joe’s waitlist. Booking needs specific approval. It cannot change Steve’s booking and has no messaging tool. When an apparently available place disappears at booking time, the tool refuses the booking; Gogo rechecks availability and uses the permitted waitlist.

A visible action log lets a person see that failure, inspect the response and intervene if necessary. The safer result comes from the context, procedure and authority configured around the same original request. Both gym sequences are scripted illustrations, not live agents accessing real accounts.

Verify the actual result

Check system state, not just the agent’s report. In the safer demo, Joe has a waitlist entry and no booking, Steve keeps his place, the gym is open, and Gogo sent no messages. “You’re waitlisted” is useful and truthful; “You’re booked” would be wrong.

Apply the same discipline to work: inspect the change, run the relevant checks, record uncertainty, and decide who accepts the outcome. A successful metric alone is not enough.

Talk to the Kohde team about a complimentary 30-minute consultation, or connect with Joe and Kohde through the resource page.