The delegation ladder.
"Agents" and "agentic workflows" get talked about like a new superpower you need special tools to unlock. Mostly, they aren't. They're a handful of moves you can already make by hand in an ordinary chat — a ladder of how much structure a task needs. The tools just automate the rungs. Here's the ladder I actually use, and the one rung almost everyone skips.
The tool doesn't give you judgement. It scales it.
Every "agentic" feature I've used turns out to be one of a few simple patterns with the manual labour taken out. The patterns are the interesting part, because you can run them by hand today, with nothing but a chat window — and once you can see them, the fancy tools stop looking like magic and start looking like a lift.
So I think of it as a ladder. Each rung is a bit more structure, a bit more autonomy handed over. The skill isn't climbing as high as possible — it's picking the lowest rung that does the job, and not skipping the one rung that keeps you honest.
The tool doesn't hand you judgement. It scales the judgement you already have — one rung at a time.
Seven rungs, from "just ask" to "loop until done".
Climb only as high as the task needs. Each rung has a no-tools version you can do in any chat right now.
Just ask
One question, one answer. Most tasks live here — and should. Reaching for machinery above this rung is the most common mistake people make once they've heard the word "agent".
By hand → A normal chat message.
Sort, then act
Let it triage before it works: which of these are X, which are Y? Then handle each bucket on its own terms instead of in one mushy pass.
By hand → Paste the list, ask it to label each item, then deal with each label.
Fan out
Ask the same question from several angles at once, then combine — breadth you'd never get from a single answer.
By hand → Open three chats (or Projects); give each a different cut; merge what comes back.
Make it argue with itself
Take the answer to a fresh chat and ask what's wrong with it. This is the rung most people skip — and the one that earns its keep.
By hand → Copy the output into a new chat: "Find the weakest points in this."
Generate, then filter
Ask for thirty, not three. Quantity first, judgement second — the good one is rarely in the first handful.
By hand → "Give me 30 options." Keep the two that survive a second look.
Run a tournament
Compare two at a time and keep the winner, rather than asking it to rank ten at once — which it does badly and confidently.
By hand → Paste A and B: "Which is stronger, and why?" Then run the winner against C.
Loop until done
Hand over the finish line, not a step count: "keep going until X", not "do these five steps". This is where real delegation starts.
By hand → "Keep refining until every claim has a source." Let it decide when it's there.
Make it argue with itself.
If you take one rung from this, take this one. An AI is a flatterer by default: ask it to check its own work in the same chat and it will mostly defend what it just said. A fresh chat has no ego in the answer — it reads the output cold and tells you where it's thin.
For anything that matters — a legal claim, a number, a plan you're about to act on — this is the cheapest insurance there is. Paste the answer into a new conversation and ask it to break it: "What's wrong here? What would a sceptic say?" Half the time it finds something. That half is the whole point.
Over-reaching and under-reaching.
Machinery for a one-liner
Wiring up a rung-6 loop for something a single question would have answered. It's slower, more brittle, and the output is usually worse. Most work is a rung-0 or rung-1 task wearing a serious face.
Shipping rung 0 on a rung-3 job
Taking the first confident answer at face value on something that actually mattered. The fix isn't a better tool — it's spending thirty seconds on rung 3 before you rely on it.
It's all one skill: delegation.
In the four-competency way I think about working with AI, this whole ladder is a single competency — delegation: deciding what to hand the machine, and how much rope to give it. The agents and workflow tools everyone's excited about are simply rungs 2–6 with the manual work removed; the judgement they need is the same judgement you'd use by hand. Which is the good news — you can practise it today, for free, in any chat.
If you want the public, business-facing version of this shift — from chatbots that answer to agents that act — it's over on the firm's site: from chatbots to agents.
More field notes
Things I've found useful and reckon are worth passing on — from tinkering with AI to the craft of working well.