Why Pólya, and why now
George Pólya wrote for math teachers, not for programmers or AI architects. That distance is exactly what makes his work useful at this moment. When the technical problem changes at speed — new models every six months, new tools every week — the discipline that survives is the one that operates at the level of thought, not syntax. Pólya described that discipline with four phases that are, once named, almost obvious: understanding the problem, devising a plan, carrying out the plan, looking back. The hard part isn't grasping the list. The hard part is not skipping phases, because the temptation is enormous.
The tech industry skips two of them. It skips the first, because "we already know what the problem is": getting the agent to work. It skips the fourth, because "it's already running": we move on to the next project. The two in the middle — devising a plan, executing — are the only ones that get done, and they get done compressed, under pressure, often simultaneously. The result is what we see: projects that work but don't transfer, teams that delivered but didn't learn, organizations that have AI in production and have no idea what to do with the next wave. Pólya described the disease before the disease existed. His book is, at this moment, more useful than many contemporary governance manuals.
The Lab applies the four phases with discipline. What follows is an honest reading of how each phase operates when the problem is delegating autonomy to an agent — and why skipping any one of them returns the organization to the same place where software engineering stood for forty years.
Understanding the problem
"It is foolish to answer a question that you do not understand. It is sad to work for an end that you do not desire. Such foolish and sad things often happen, in school and out of it, but the teacher should try to prevent them from happening in his class."
Most AI projects in organizations are built on a misunderstood problem. The question teams believe they are solving is how to make the agent work. The question that actually matters — and that the organization will discover later, almost always after an incident — is what responsibility we are delegating, to whom, with what reversibility. The first is a technical problem. The second is an organizational one. Confusing them is not a detail: it is the source from which almost all subsequent governance problems derive.
Pólya insisted that understanding the problem means answering, with precision, three elementary questions: what is the unknown? what are the data? what is the condition that links them? Applied to the problem of delegating to an agent, the answers are not obvious and tend to be ignored. The unknown is not "the agent's output": it is who is accountable for the decision when there are consequences. The data are not "the training dataset": they are the scope of delegation, the cost of error, the reversibility of the action, the maturity of the human team that supervises. The condition that links them is not technical: it is contractual. What an organization delegates defines what that organization can be held accountable for.
That's why the Lab's first move when faced with an AI project is not to propose architecture. It is to ask — with the insistence Pólya recommended — what concrete problem we want to solve, in what domain, with what consequences if it fails. That dialogue, austere and sometimes uncomfortable, is the phase most projects skip. When skipped, everything that comes afterward is built on sand.
Devising a plan
"Here is a problem related to yours and solved before. Could you use it? Could you use its result? Could you use its method?"
This is the most underrated question in the book. Pólya argues that when faced with a new problem, the first intellectual operation is not to look for an original solution — it is to look for a neighboring problem that has already been solved and ask whether its method transfers. That move is the opposite of the contemporary instinct, which celebrates the original solution as a sign of talent. Pólya, who was a mathematician and therefore understood something about originality, defended the opposite: the mature solver recognizes the neighboring problem before lifting the pencil, and that's why they solve faster and better than the novice who insists on starting from scratch.
Applied to the problem of delegating autonomy to an agent, Pólya's question has a clear and almost always ignored answer: the problem of delegating autonomy is not new. Human organizations solved versions of this problem for centuries. Corporate governance solved it for boards and CEOs. Accounting audit solved it for financial transactions. RBAC solved it for software infrastructure. The law of mandate solved it for proxies and representatives. Each of those domains developed a vocabulary for stating what is delegated, in what scope, with what supervision, with what trace. What is new is not the problem. What is new is the substrate — a probabilistic agent — and, as Pólya warned, the method of the neighboring problem almost always transfers if one takes the trouble to adapt it.
That's the intellectual operation of the Lab's first essay: ARL does not invent the autonomy contract. It takes it from RBAC, adapts it to the agent's substrate, articulates it in five operational levels. It is not innovation; it is rigorous continuity. Novelty would be the error. And yet, the industry's reflex is to invent new frameworks every year, without acknowledging that the problem already has a century of resolutive tradition behind it. Pólya would have called that pretension a lack of craft.
Neighboring problems · transferable method
→ delegation to board and CEO
→ transactions at scale
→ delegated technical action
→ proxy and representative
What the Lab does, phase by phase, is translate these neighboring methods to the new domain. ARL translates RBAC and the law of mandate to the agentic substrate. Threshold-based traceability translates accounting audit. Digital Symbiosis translates the principle of corporate governance according to which the quality of the decision depends not only on whoever decides, but on the collaborative structure within which they decide. None of this is invention. All of it is contemporary application of methods Pólya would have recognized as a legitimate move of the craft.
Carrying out the plan
"Carrying out your plan of the solution, check each step. Can you see clearly that the step is correct? Can you prove that it is correct?"
The third phase is the one the industry best understands and worst practices. Best understands because execution is what gets measured, what gets delivered, what gets billed. Worst practices because Pólya warns of something almost always omitted: each step must be verified while it is being made, not afterward. Subsequent verification is always more expensive and often impossible — the problem is discovered when the system is already in production, the data has already migrated, the decision has already been made. Pólya knew this and stated it without drama: checking step by step is the cheap way of not having to redo the entire work.
The Lab's frameworks — ARL, Digital Symbiosis, Cognitive Footprint — are the instruments that materialize the plan in the execution phase. ARL fixes the scope of delegation. Digital Symbiosis fixes the quality of the collaboration. Cognitive Footprint fixes the admissible cost of that collaboration. Each operates at its own layer, and the three operate together over the same contract. But an instrument is only that: an instrument. It does not replace the verification Pólya called for. Verification has its own apparatus — delegation logging, human escalation thresholds, drift monitoring — developed in the essay on tracing agents at scale. Here it is enough to name it: execution without intermediate verification produces the same errors as execution without a plan.
What is worth underlining in this phase, and lost when treated only as implementation, is that Pólya's discipline in execution is not bureaucracy. It is the opposite of bureaucracy: it is the economy of work well done. Verifying while executing does not add time to the project — it saves it, because it prevents the rework that is the most invisible and most expensive cost category in any technical system. The organizations that best implement AI are not the ones that deliver fastest: they are the ones that verify best while delivering. That is a Pólya intuition applied with eighty years of delay.
Looking back
"Even fairly good students, when they have obtained the solution of the problem and written down neatly the argument, shut their books and look for something else. Doing so, they miss an important and instructive phase of the work."
This is the phase Pólya defended most insistently and the one almost all his readers forget. He called it looking back, and argued that it is the phase where the solver builds themselves. Without the fourth phase, the problem was solved but the solver did not learn. Each future problem is faced from scratch. Capability does not accumulate. That, transposed from the classroom to the technical project, is exactly what has happened to software engineering for forty years: every company, every team, every project reinventing the wheel — because the fourth phase was never done at the level of sustained discipline.
Pólya asked two questions when looking back. First: is the solution correct? That question is relatively easy; most teams ask it. Second: can the method be used to solve other problems? Almost no one asks that one. And it is the question that separates the team that delivered a project from the team that built capability. It is the difference between the team that got an agent running for one case and the team that discovered a delegation pattern that serves the next ten. The first is a service. The second is an organizational asset.
The Lab exists, in good measure, because someone has to sustain the fourth phase at the level of discipline, not project. Each AI implementation at a client is a particular case; the method distilled from those cases — what delegation pattern worked in what context, what thresholds were correct, what supervisions proved effective — is what becomes an asset. ARL, Digital Symbiosis, Cognitive Footprint did not arise from theory. They arose from the fourth phase applied with discipline over real projects. They are distillations, not speculations. That difference is what justifies their existence.
It's worth saying frankly: the fourth phase is not profitable in the short horizon. Looking back and examining the method takes time no client pays for directly. It is an investment made against the next project, not the current one. Most consultancies don't do it because their business model doesn't reward it. The Lab exists as the organizational answer to that problem: it is the internal institution that sustains the phase that the business, left to itself, would neglect. It is not a luxury. It is the structure that makes learning accumulate instead of evaporating with each delivery.
The method is the asset
If the Lab's frameworks were understood as finished products, their value would be limited. Yet another framework in an industry that produces frameworks as disposable goods. The intent is different. The frameworks are the visible expression of an invisible discipline: the method applied to solving AI governance problems, refined case by case, transferred between projects, sustained institutionally. What matters is not ARL or Symbiosis or Footprint taken as artifacts. What matters is the method that produced them and that will produce the next ones when the substrate changes.
Pólya understood this in 1945. He formulated it for math teachers, but he formulated it as a general truth about how capability gets built: the solver builds themselves through the work, and the phase that builds the most is the fourth — the one done when there is no more urgency, when the problem is already solved, when almost no one is watching. That phase, which seems optional, is the only one that distinguishes the team that delivered from the team that learned. And in the long run — longer than quarterly metrics measure — it is the only one that distinguishes the organization that will be ready for the next problem from the organization that will be caught off guard.
The era of agents is going to produce many next problems. New models, new capabilities, new risks, new regulations. Each wave will test the organization's ability to respond. The one with method will respond. The one with only past projects will have to relearn each time. Pólya wrote the manual of the first one. The Lab is a contemporary attempt to apply it with the seriousness it deserves.