ARTICLE · CONTROLS FOR AI WORK · PART 3 OF 6 · Sequencing · ladder logic

Playbooks: steps, gates and “done when”

A skill tells the AI agent how to do one job. A playbook tells it which job comes next, and what has to be true before it moves on. On a machine, that’s sequencing: steps in order, each one waiting on its conditions. It’s the part of controls work people underestimate, and it’s the part that keeps a long job from wandering off.

The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?

What the AI field has found

A widely read industry guide, “Building Effective Agents,” separates two kinds of AI systems. A workflow follows a path written in advance; AI does the initial analysis at each step. An agent decides its own steps as it goes. Their advice is to find the simplest solution that works and add complexity only when it’s needed. Agents cost more and their errors compound.

One of the workflow patterns they describe is a chain of steps where you “add programmatic checks” between them, a gate: the output of one step is checked before the next step starts. If the check fails, the chain doesn’t continue.

That’s a playbook. A fixed order for work you do again and again, with a check at each handoff.

From the shop floor: the scan cycle and the rung

Sequencing. An automated line runs in steps. The part is in the nest. The clamp closes. The clamp-closed sensor is on. Then, and only then, the robot moves. Every step waits for its conditions, and no step starts because the previous one probably finished.

Ladder logic. In ladder logic, an output turns on only when every condition on its rung is true. Not most of them. All of them.

The scan cycle. A PLC runs in a loop: read the inputs, run the logic, set the outputs, repeat, many times a second. It checks as it goes, every pass.

Lately there’s a lot of talk online about “looping” AI agents: let the AI agent run over and over until it decides it’s finished. Anyone who’s programmed a PLC has lived in a loop for years. The loop was never the smart part. The smart part is what gets checked each time around, what stops it, and what happens when it’s wrong. A lot of the looping advice runs it the other way: let the AI agent go for a long time, then check at the end. That’s like running a line for a full shift and inspecting only the last part. Check each piece as you go, and stop as soon as the path is wrong.

How it carries over: “done when”

In a playbook, every step ends with a “done when”: the conditions that have to hold before the work moves on. Here’s the one from the A.N. Creation Office’s design playbook, which takes an inventor’s design from first concept to a frozen design ready for tooling:

DONE WHEN: The design is frozen at one named revision, which needs all five of these:

  • every compliance matrix line is met or put off on purpose, with the reason and who agreed;
  • at least one validation with real use is done;
  • the remaining risks are accepted by the inventor;
  • the cost estimate and timeline still fit the charter, or the inventor accepted the trade;
  • the drawings, parts list and material specs are released at one revision.

That’s a rung of ladder logic. Five conditions in a row, all true, then the output fires: the tooling playbook starts. Four out of five doesn’t freeze the design, and the AI agent doesn’t get to decide that it’s close enough.

Notice who decides on two of those conditions: the inventor. A gate isn’t only a check the AI agent runs on itself. Some conditions are a person’s yes.

Keep the state outside the chat

A machine always knows which step it’s on, because the program keeps track. A conversation with an AI agent forgets. Memory gets cleared, conversations get long, sessions end.

So a playbook writes its current step down in a file: what’s done, what’s next, what’s waiting on whom. Every new session starts by reading that file, not by trusting what the AI agent remembers. Published engineering guidance recommends the same thing for long jobs: have the agent keep notes outside the conversation and read them back. It’s the same reason a plant keeps a traveler with the parts instead of relying on whoever ran the last operation.

When not to use a playbook

Playbooks are for work you do again and again, where the order matters. Taking a design to freeze. Preparing a funding packet. Onboarding a new customer.

They’re the wrong tool for exploring. When you’re brainstorming or asking for an opinion (Part 1), a fixed sequence gets in the way. Give the AI agent room there, then bring what you decide back into the playbook.

Try this

Pick one job you do with an AI agent more than once. Write it as three to six steps. Under each step, write “done when” and two or three conditions you could check. Next time, have the AI agent work one step at a time and check the conditions before it moves on. Keep the list of what’s done in a file, not in the chat.

Sources

These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.