The Shape Coding Agents Converge On
Every serious coding agent is converging on one shape. The same hard limits force the same parts, whatever brand sits on the front. Build the ring around the model, because the ring is where the value stays.
Every idea in this guide showed one side of the same machine. The middle of that machine is the part you swap out.
Convergence model
The whole series sits in one picture. The block in the middle is the model, and it slots in and out. One model today, a cheaper one next, then the one after that. Each takes the same slot.
The workshop around it holds still. The notebook, the window, the traffic light, the tiny desk and the safety net stay where they are. The limits build that workshop for you, and swapping the brain does not touch it.
The brain gets the headlines. The workshop is the product.
I started seeing that machine in agents other people built. The parts sat in the same places. Step the figure below through the ring and watch each part light up in turn.
01One architecture, said in one breath
A coding agent is a loop around a model that forgets everything between calls. The loop hands it a set of tools, and each tool is a promise about what comes back.
A traffic light stands in front of those tools, and the light can be answered from anywhere. Behind all of it, the window is a budget the loop defends.
Four parts, and none of them is the model. A limit put each one there.
02Why they converge: same constraints, same shape
This part surprised me. Each limit below is a wall you hit in real use. The part beside it is what you build against that wall.
- The model forgets between calls. The loop refills the notebook every turn and keeps each request paired with its result. That is the loop and the worldview.
- The window is small and every word costs money. Context and the prompt economy answer it. You set a budget and reuse what you can, and the order you write in sets the bill.
- The model will sometimes ask to do unsafe things. The light goes in front of its hands, and you scope what it touches before it starts. That is tools as promises and permission as a real gate.
- Reality breaks things mid-run. You recover, or you stop with a named ending and a reason. Resilience is that pair.
- Real work runs longer than one chat and wider than one window. Memory and subagents split that load. Shelves hold what outlives the session, and a tiny desk takes the scoped jobs, cheap move first.
- Nobody can guess every use ahead of time. So you leave doors open for other people to snap parts on. The pluggable layers are those doors.
Start with a real model and real people using it, and the limits walk you here. Nobody had to copy. The shape is what is left standing once the problem is done with you.
03The model is becoming a component
Follow the limits down and you arrive at the swap. You pull the model out, drop a better one in, and the ring around it does not move. That is an engine change, and the car stays the same car. Every limit in the list above still forces its part, whichever model sits at the center.
Where the work lives
The failing-well idea already priced this split. The work that makes an agent safe to trust on a plain Tuesday lives in the plumbing. Press the swap in the figure below and see which half of the picture changes.
04What the shape tells you to build
The shape hands you a build order. Decide what the loop carries between turns, and how the worldview goes back the same way each time. Write down what each tool touches, and who grants the permission. Set the window budget, and decide what gets loaded first. Name what resilience does with a failed step, what memory keeps, and what you drop on purpose. Scope each job you send to a tiny desk, check what comes back, and leave studs for other people.
Get those right and a so-so model turns into a useful agent. Get them wrong and a strong model is a demo that falls over. Press the controls below and the rule sorts a bolted-on part from a real gap.
The honest caution
This shape is unfinished. The last idea named the half-built corners: keeping work alive, driving it from far away, many agents working together. Those corners are the parts still arriving, and the next limits will add more. A part copied because the picture has one carries no weight, and it bills you anyway.
Decision
When you build an agent, the workshop comes before the prompt. Point at each part of it and say out loud the limit that forces it to exist. Get through all of them and you understand your own system.
The model in the middle keeps getting better, and someone else pays for that. The workshop is yours. Cut the part you cannot explain, and build the limit that has no part yet. That is the audit I leave the series on, builder to builder.