Controlled experiments, benchmarks, and reference-design notes from the Poesis engineering team.
How these ideas fit together
These notes are not independent essays. A handful of foundational pieces, tagged Premise below, state the premises; the rest are what those premises imply once applied to a specific question.
The standard indictment of language models — that they reason probabilistically — describes human intelligence just as well. Humanity never remedied this by making minds deterministic; it built methodologies and institutions around them. Generative AI automates our reasoning. It should inherit our discipline as well.
Applied in: agentic harness · agentic SAFe
Poesis takes a single position on the challenges of operational generative AI: concerns are owned by stakeholders and carried by governance, never by the model. The model proposes; the stack and the human dispose. This article states that position once and shows how each challenge resolves under it — the discipline the market is beginning to name operational AI governance.
Applied in: agentic AI governance
Autopoiesis names the capacity of a system to produce the components that produce it. Generative AI has made self-production cheap for organizations and software alike — what it has not made is the conserved organization that keeps self-production from becoming dissolution. That is the gap Poesis is named for — and the missing foundation of the autonomous enterprise.
Applied in: autonomous enterprise · agentic harness
Poesis and GSM are definition-centric because no description is definition-free. Definition makes observation intelligible; description supplies evidence from realization; that evidence can then inform the next definition.
Applied in: governed world model · autonomous enterprise · digital twin
Retrieval answers which passages resemble a prompt. Enterprise assistants usually need something else: which definition is in effect, which policy binds this scope, and which action is permitted. Where a domain is modeled, governed context can replace the retrieval layer rather than supplement it — the layer context engineering is still missing.
Applied in: agentic AI governance (conduct vs. consequence)
Classical systemics describes systems after the fact. GSM inverts that — it defines the system, and the definition generates and governs the running thing.
Applied in: spec-driven development · digital twin
Regulators, consultancies, and research journals converged this year on the same conclusion: as AI agents gain authority, governance becomes the primary constraint. We think that consensus is right, and this note develops the question it opens — how governance can operate at the tempo of what it governs — through an old cybernetic principle and the four design commitments we arrived at when we tried to honor it.
Of everything the agentic era is reaching for, a method may matter most. SAFe already solved the problem agent builders are rediscovering — how to get reliable delivery out of many bounded, fallible workers — which is why we chose to run it with agents rather than invent a new coordination scheme. This note develops the reasoning and shares what running SAFe agentically actually looks like.
The digital twin is one of engineering's most trusted patterns: a live model, continuously fed from the real asset, that you act on before touching reality. Extending it to the organization itself is the right ambition — and it inherits a hard question the physical version never had to ask: what are the sensors, and what is the physics? We share what we learned building one, and why the answer begins in IT.
The autonomous enterprise is becoming a serious category, built on a genuinely useful distinction: automation executes procedures, autonomy decides. This note brings a piece of systems theory to the conversation — autonomy has a precise meaning in the biology of cognition, and taking it seriously suggests what the category’s foundation has to be: an explicit, governed definition of the enterprise itself.
A learned model can retrieve and reason over what an enterprise has decided. It cannot, by learning alone, make a definition current, an obligation binding, or an action permitted. Enterprises therefore need a complementary model of their definitional and institutional world.
The agentic ecosystem has converged on the word harness from several directions at once — and when practitioners converge on a word like that, they have usually found something real that needs a name. This note develops the concept: where it comes from, what it has to do, and what we learned building one — including the one property that turned out to carry all the others.
We watched spec-driven development emerge with a sense of recognition: it is the same inversion we bet on — when implementation is generated, the definition becomes the artifact of value. This note explores the question the movement will meet next, one every practitioner already knows from documentation: what keeps the spec true? We share the answer we converged on — give the spec a lifecycle — and how it closes drift in both directions.
The standard indictment of language models — that they reason probabilistically — describes human intelligence just as well. Humanity never remedied this by making minds deterministic; it built methodologies and institutions around them. Generative AI automates our reasoning. It should inherit our discipline as well.
Poesis takes a single position on the challenges of operational generative AI: concerns are owned by stakeholders and carried by governance, never by the model. The model proposes; the stack and the human dispose. This article states that position once and shows how each challenge resolves under it — the discipline the market is beginning to name operational AI governance.
Autopoiesis names the capacity of a system to produce the components that produce it. Generative AI has made self-production cheap for organizations and software alike — what it has not made is the conserved organization that keeps self-production from becoming dissolution. That is the gap Poesis is named for — and the missing foundation of the autonomous enterprise.
Poesis and GSM are definition-centric because no description is definition-free. Definition makes observation intelligible; description supplies evidence from realization; that evidence can then inform the next definition.
Language models are asked to infer organizational reality, authority, and permission from prose on every request. Poesis changes the environment: models propose against governed definitions, while deterministic and human governance decides what may become action. Smaller models may benefit most — a hypothesis the architecture makes testable.
Retrieval answers which passages resemble a prompt. Enterprise assistants usually need something else: which definition is in effect, which policy binds this scope, and which action is permitted. Where a domain is modeled, governed context can replace the retrieval layer rather than supplement it — the layer context engineering is still missing.
A framework-independent reference model for turning regulatory obligations into governed definitions, evaluable Norms, and continuous evidence without flattening each framework into a generic control list.
Open-ended research is not a workflow to accelerate but an institution to constitute. The recurring failures of agentic research — judgment, resources, feedback, backtracking, compliance — are one failure seen five ways: institutional work performed without institutional semantics.
OpenTelemetry gave the RUN layer a neutral standard. The definitions those runtimes enforce — the THINK layer — still have none. Here is why that matters.
Classical systemics describes systems after the fact. GSM inverts that — it defines the system, and the definition generates and governs the running thing.
Poesis