Version 13.91: Why Encoding Judgment Now Is the Only Move

The Corner Column Problem

Somewhere in a structural engineering practice, a senior engineer knows what to do when two columns land too close together at a grid intersection. She has known it for twenty years. It is a few seconds of judgment — dissolve one column, favour the corner, extend the beam slightly. It happens on every project. It has never been written down.

That heuristic, and several thousand like it, constitute the true intellectual capital of a technical design firm. Some live in BIM templates, or standards documents, but mainly they live in the heads of people — and the heuristics are manually, and imperfectly, applied on every project.

What Agentification Actually Means

I'm using the term loosely: for this essay an agent is software that executes encoded judgment — probabilistic or deterministic, AI model or API, or some combination. What matters is that the judgment is encoded rather than manually performed.

When the structural team formalises the corner column heuristic — a few dozen lines of logic, edge-checked and tested — they have done something categorically different from drawing a column. If done well, they have removed that problem from the class of problems the firm will ever solve again.

Version 0.2 handles corners. By version 13.91, the team has not thought about corners in five years. They are working on other stuff.

The Compounding Logic

Each encoded heuristic is embodied IP. It does not leave when the senior engineer retires. It does not get prosecuted from first principles by the graduate on the next project. It accumulates — structural logic, services coordination, façade parameterisation, landscape placement — each agent incrementally more sophisticated than its predecessor, each iteration building on a base that never resets.

The economics of this compound are extraordinary, because the second judgment has a marginal cost to deliver of about $0. A firm that begins encoding in 2025 and a firm that begins in 2028 are not three years apart. The 2025 firm has three years of compounded specificity — edge cases resolved, heuristics refined, agents coordinating with one another across disciplines. The 2028 firm starts from nothing, against a competitor on version 13.91, with SaaS-style economics.

But Can You Actually Appify Consulting?

I'm not naive to the insurance implications. I don't believe you can just "appify" consulting work. But this appification happens informally in firms all the time. Structural rubrics are encoded in spreadsheets and memory and passed around.

Software is getting increasingly sophisticated, and the economics are not going to change or disappear. Given the economics will stay the same and the direction of travel is clear, the insurance, legal, and professional indemnity implications are under enormous pressure to change. They will probably change.

What This Means for Developers

For the property developer, the implications are large. The consulting stack — from site discovery and planning, through structural, services, façade, landscape, core planning — has traditionally been a necessary cost of translating a developer's brief into a buildable design. That stack exists because the knowledge required to produce that design has been distributed across specialists who each require significant coordination overhead.

As agents encode and automate increasing proportions of technical design work, the stack does not disappear, but its ratio of computer to human shifts — more output becoming algorithmic, less requiring bespoke reconstruction by human teams. The developer's relationship to the consulting stack is entering a period of significant renegotiation.

Legacy Software's Structural Exposure

Legacy software vendors occupy a deteriorating position in this shift. Their moat was constructed on the assumption that geometry manipulation was complex enough to require sophisticated interfaces and years of training. That assumption made sense when design work was primarily manually inserting remembered heuristics and judgment into the model by hand.

The interface complexity that justified Autodesk's product investment is precisely what algorithms make unnecessary. A developer changing a structural grid parameter does not need Revit — or array tools and ribbon commands. The agent places and manipulates the columns. The assets built to support manual editing — the interfaces, the training ecosystems, the seat licence models — are stranded against a workflow that has no use for them.

That is my guess as to why Autodesk and Graphisoft are moving upstream.

Where Giraffe Sits

Giraffe has positioned itself at the intersection of these forces. Design, capital, cities and automation. Buildings remain systems of systems — structural, services, façade, landscape, vertical transport — and those systems must coordinate against shared geometry and shared data.

What is changing is the nature of the work being done within each system. Giraffe operates on open data formats, agnostic to whether a given output is produced by a probabilistic AI agent, a deterministic API, or some combination of both. The geometry is always available to the next agent in the chain.

This video shows agentic workflows operating on a simple, declarative building mass. The building mass is enough for the finance and the planning team to do their work. It is also enough for the agents to create a detailed model based on some heuristics and assumptions.

These are not complete, detailed, or even good assumptions. They are mine — but you can see very clearly how each discipline, staying in their lane, would be able to build increasingly complex, robust and powerful models that solved buildings.

Who Should Build the Agents?

We can see some software firms beginning to build their own agents. At Giraffe we are resisting this because we think that each discipline should build its own agents. Structural engineers should build structural agents. There is a depth of domain expertise that it's difficult for a software vendor to obtain.

Engineers do face an innovator's dilemma, obviously — an agentic workflow would force hard questions about their existing business models. The purchase of Consigli by AECOM and its description as existential shows how boards are grappling with this dynamic.

---

For the technical design firm, the strategic choice is immediate. A firm's accumulated heuristic knowledge is either encoded into agents — where it compounds, scales, and becomes a structural advantage — or it remains in the heads of different individuals, where it is scattered, slow to transfer, and increasingly uncompetitive against firms that have made the other choice.