Agent topologies and the shape of the work

Making that second move takes two things. He has to understand how a system of agents can be shaped, which is the rest of this piece. He also has to see that those shapes are not foreign. They are the structures he already knows from running an organization full of people. That second idea is the important one. I would ask that you keep it in mind as you read.

Now read the table again. Once you look at it the right way, the design questions mostly answer themselves.

The steps and their order tell you how many agents you need and what blocks what. Each step marks where an agent could go, not where one must, the shapes refine that count later. The order matters, because one step's output becomes the next one's input, and the call's own input feeds the first step while its result falls out of the last. You cannot judge a dispute until you have pulled the charges, which is a hard dependency the design has to respect. The knowledge column tells you what to hand each agent. Some of it is fixed, like the refund policy, which is the same on every call and can be cached once and reused (caching gets its own section below). Some of it is a judgment, like whether this is really a duplicate, and that has to stay with a careful agent or a person. The data column tells you where the wall belongs, the privacy boundary you will meet properly in the features section. The sensitive systems are the ones that sit inside that boundary, reached only through what that section calls the internal agent. The two steps marked as judgment, the ones that run on reasoning instead of a rule, are the ones a person keeps a hand on.

None of that needed a new skill. It needed the leader to write down what they already knew, in a form the design could use. That is the job, turning what you already know into a form the system can run on. It is the input to every decision that follows.

The language model is only the reasoning in the middle. Triage, judging the dispute, sizing the refund, choosing the words. That is where every shape here lives, and where the cost and the latency and the risk collect. Everything around it is cheaper, faster, and more reliable for the plain reason that it is not a language model.

Reaching for an LLM to look up a balance or play a hold message is the same mistake as spawning fifty agents for a one-line question. The best systems put the model exactly where judgment is needed and use plain code everywhere else. That sharpens the first of the four questions, because Q1 has two layers and it pays to name both. The first layer is which work gets an agent at all, which the “what stays human” section takes up later. The second layer, this one, is that within the work you do hand over, only some steps need an LLM, and the rest should be plain code.

Everything in this piece, the shapes, the routing, the caching, lives inside that one red box. It is one slice of a larger system, and knowing where that slice starts and stops is its own kind of expertise.

The organization already knows this one. A rep does not relearn the refund policy on every call. It is in the binder, pinned to the monitor, already in their head. What they pull up fresh is this caller's account. Caching is that binder. The standing knowledge that does not need re-reading, kept close so the work stays cheap.

It changes what every shape costs to run. The pyramid looks expensive until you remember the policy its reviewer checks against is cached, so each review costs far less than the gauge suggests. Caching does not change which shape you pick, only what it costs to run.

One caution: “deflection” and “resolution” are not standardized terms, every vendor defines them differently, so ask for the definition before you accept the metric.

Which is why the leader we met at the summit is not starting from zero. He has run these structures with people. He knows a delegating manager is fast but misses things, that a review chain is slow but safe, that the back office exists to keep secrets in. Henry Mintzberg, who spent his career mapping how organizations are built, catalogued the shapes organizations take, the simple owner-run shop, the machine bureaucracy that runs on process, the professional firm of experts who manage themselves, the divisional structure, and the loose adhocracy built for one-off problems. These shapes are the same family, drawn for agents instead of people. The leader already knows which fits which work, in his gut, because he has lived inside them.

Agentic design asks the leader to apply exactly that judgment to a system that runs at the speed of software. That should feel like familiar ground. There is rarely one best structure. Several can work, each at a different cost, and the skill is matching one to the work in front of you. Organization theorists have a name for this, equifinality (Gresov and Drazin gave it its canonical treatment in 1997), the idea that different designs can reach the same end by different routes. So the question that looked new, how many agents and what shape, is the organization-design question the leader has answered a hundred times. Those years running the operation are exactly the qualification this work asks for.

This is also where the reps fit. When the routine calls collapse into agents, the hard and human ones do not disappear, they concentrate. The reps stop taking those calls themselves. They watch the agents, decide which shape to run for a given call, approve the actions that cannot be undone, and personally handle the hardest, most sensitive ones the system escalates. That is a harder job than the one before it, and there are still plenty of calls that need a person.

A 2026 Anthropic study, “AI Organizations are More Effective but Less Aligned than Individual Agents,” found that groups of agents outperform a single agent on the task while adhering less to the safety guidelines they were given, a diffusion of responsibility in which decisions pass between agents until nobody, including the human receiving the result, is accountable for it. More agents may require more human accountability infrastructure, not less, which is exactly why the choice between in-the-loop and on-the-loop is a design decision and not a detail.

The diagonal is the lesson. Move from the bottom left to the top right and the person goes from out of the loop, to signing off on the agent’s work, to leading it while the agent assists. That is the automate to augment to human spectrum, and it maps straight onto the shapes you already have. The familiar rule for what to hand a machine is the three D’s, the dull, the dirty, and the dangerous. What stays with a person is the rest, the work that takes judgment or someone to answer for it. Today the top right corner stays human. The forecast at the end is where I ask what moves.

This starts with the customer. They are talking to one agent that holds nothing sensitive. Behind the wall the work fans out in parallel, and each branch runs the shape that fits it. A hub and spoke gathers the account details, because the lookups do not depend on each other. A pyramid gives the money back, because real funds move and any tier can send the refund back before a dollar does. A pipeline handles the after-call wrap up, because those steps have to happen in order. None of that reaches the customer. That is the idea in one picture. You compose the shapes, you do not pick just one, and it can be this involved or as simple as a single agent, depending on what the work needs.