Building a Self-defining Agent Organization

Part of the write-ups behind my Build Stuff 2026 talk, Control Your Fate at the Metal, Vilnius, 2 December.

Most work on agents is about the agents. Better planners, better tools, better prompts, better memory. The organization around them, how goals are set, who owns what, how work is handed off, when a human is asked, is designed once by an engineer and frozen.

That is backwards. Agents are cheap and disposable. The organization is what achieves goals. So the thing to build is not a set of agents but an organization that evolves itself to achieve its goals, inside a structure that humans govern at the most strategic points.

Goals are the reason the organization exists

An organization is not a collection of capabilities. It is a set of goals and whatever it takes to achieve them. Revenue closed. A product shipped. A market entered. Each goal has an owner and criteria that are met in the world or not met at all.

The goals form a graph. A company outcome at the top, function goals beneath it, each supporting the one above. The graph exists so that success can be traced: from a company result, down through the goals that produced it, to the work that moved each one. Any task that cannot be traced to a real-world outcome through that graph is not work. It is motion.

Everything else is instrumental. Roles, procedures, tasks, documents, reviews, the agents themselves: all of it exists to move a goal, and none of it is value on its own. An organization that is busy and achieves nothing has failed, however impressive its activity looks.

So the first question is never "which agents do we need?" It is "which real-world outcomes, and what would it take to produce them?" The answer to the second half is allowed to change.

The organization evolves itself

Human organizations learn by changing their structure. They hire for roles they did not know they needed, write down a procedure after the third time something went wrong, split a team when it owns too much. Nobody files a ticket with engineering to do this. The organization rewrites itself.

An AI organization can do the same if it is written in a language its agents can read and write. Define the organization as documents: what each function is for, what it owns, how work is done, who may do what. The agents read those documents to know who they are. When a goal needs a capability that does not exist, an agent writes the description of the specialist that will provide it. When a procedure fails, the organization proposes a better one. When a goal has no owner, the organization proposes the function that should own it.

The recursion runs all the way up. Specialists are composed for a task and discarded. Functions persist, own goals, and commission specialists. The organization as a whole reviews its own performance against its goals and changes the functions and procedures that fell short. Each layer is a document the layer above can rewrite.

Learning, in this picture, is not remembering what happened. It is changing the definition so the next goal is achieved differently. A lesson that stays in a notes file changes nothing. A lesson that rewrites a procedure changes every cycle after it.

The structure humans govern

An organization that can rewrite itself needs a structure it cannot rewrite, and that structure is where the humans are.

                    HUMANS
                      |
   set direction  .  define what success means  .  decide what may leave the building
   approve spend  .  approve what the organization may become
                      |
                      v
   +------------------------------------------------------------------+
   |  CONSTITUTION  .  PERMISSIONS  .  HUMAN GATES  .  BUDGETS         |   fixed
   +------------------------------------------------------------------+
                      |
                      v
   functions . roles . procedures . knowledge . tasks . artifacts          evolving
                      |
                      v
                    GOALS ACHIEVED

The fixed layer is small. A constitution: what the organization values and may never do. Permissions: what each kind of work may touch, under least privilege. Human gates: the decisions that stop the organization until a person answers. Budgets: how much time and effort a goal may consume before it stops and asks. The organization may not change any of these, however inconvenient they become. Humans may, in the same documents, in the same language.

Everything beneath that layer is the organization's to evolve.

At the most strategic points, not every step

Where humans sit in this structure matters as much as that they sit in it. Put a human on every action and you have built a tool that needs an operator. Remove the human and you have built something nobody governs.

The strategic points are the ones where human intent carries information the organization cannot derive: which direction the company takes, what a goal's success means, what may be published, spent, deployed or committed to, and what the organization may become. Those are gates. The organization stops at them and asks, in one page: the decision, the alternatives, the evidence, a recommendation, and why a human is needed.

Decision: Choose the primary market strategy for next quarter.
A: Agency expansion   B: Direct enterprise   C: Hybrid
Recommendation: C
Why human input is required: the choice depends on risk tolerance and
the direction you want the company to take. Nothing the organization
can read answers that.

Everything that is not a gate, the organization owns. Research, drafting, critique, planning, and the judgment calls inside them are never escalated. A human who is asked to confirm work the organization is competent to do is being wasted, and the organization that asks is failing at its job.

The human's answers are binding, cheap to give, and recorded. The organization re-plans from them. Over time the human spends less attention and the organization achieves more, which is the only direction this relationship should move.

How you know

Not by output. Plans, tasks, artifacts, reviews, cycle counts and token totals are exhaust. The dashboard that makes an agent system look busy says nothing about whether it is working.

By outcomes in the world, traced back through the goals that produced them. Revenue that arrived. Accounts that were created. A product that shipped. Each outcome traces down the goal graph to the functions that owned the goals and the work that moved them, and each change the organization made to itself is judged by whether an outcome moved because of it. Success that cannot be traced to the world is not success.

Within the structure, with human attention spent only at the strategic points. Give the organization a goal its current shape cannot achieve. Watch whether it reshapes itself, achieves the goal, and achieves the next one faster because of what it changed. Watch how often it asked, and whether every question was one only a human could answer.

If the tenth goal is achieved because of what the organization learned on the first nine, and the human was consulted only where it mattered, the organization is evolving. If not, you built a collection of agents with extra steps.

Build Stuff 2026

The organization in this article runs on infrastructure we chose to own rather than rent: our own loop, our own state in Git, our own gates. Why we go that far down the stack is the argument of my talk at Build Stuff in Vilnius this December, the 15th edition, under the theme The Age of Agency.

Control Your Fate at the Metal
Wednesday 2 December, 15:00. 45 minutes.

If you're coming, say hello. If you're not yet, 2026SPEAKER_20 takes 20% off a ticket.

#BuildStuff15 #buildstuffconf #AgeOfAgency @buildstuffconf https://buildstuff.events

I'm Kent Alstad. CEO and innovator at Cloud.Army. I build KandiForge, a toolchain for running software projects with AI agents under an evidence-based process. If you run an organization of agents rather than a set of them, I'd like to hear what you let it change about itself.