I first came across Giraffe in 2020 or 2021, during my Cerulean Labs days. Met Rob Asher and the team properly at NXT BLD 2025. They were not at this year's event, but I have been catching up with Rob in conversation recently and reading back through their public material in some depth.
This is the post promised at the end of The Word "Agent" Is Doing Too Much Work - because Giraffe is the clearest working example I have seen of the pattern I was reaching for in that piece.
A Genuine BIM 2.0 Player, With a Broader Frame
Giraffe is, by every behavioural test you might apply, a BIM 2.0 player. Real-time parametric model. API-accessible data. Modular agents and workflows. Hybrid AI and deterministic execution. Computational designers publishing reusable flows. The capabilities sit alongside Arcol, Snaptrude, Motif, Qonic, and Hypar - and in several dimensions are more mature than the cohort average.
Their own commercial frame is wider than AEC alone. As Rob has written publicly, AEC is too small a frame. AEC fees are roughly 10% of construction cost, which is roughly 30% of asset value, so even a generous reading of the AEC software market is less than 1% of the value it sits inside. Giraffe's primary buyers are developers, asset owners, capital managers, and governments. AEC professionals are part of the audience, not the whole of it.
Both descriptions can be true at the same time, and I think they are. The wider commercial frame is where the buying power lives, and it shapes what they choose to build. The product itself, looking year-on-year, is one of the most mature BIM 2.0 platforms in production right now. Anyone in the AEC conversation tracking this category should be watching them - regardless of how the company labels itself.
The AI → Code → Tool Loop
This is the part of the agentic conversation that should be getting more attention than it is.
Martyn Day's line at NXT BLD this year was everyone is a developer. Tools like That Open Company, xeokit, and the open-source IFC stack have lowered the floor on building BIM-adjacent software dramatically. AI has lowered it further. Anyone with a clear specification and a few days can now produce something that would have required a small engineering team five years ago.
However, the follow-up is that the platform-grade work - data model, geometry kernel, parametric system, federation logic, standards interoperability, the multi-year trust building that an authoring environment requires - is still not realistic for an individual firm to take on, even with AI assistance. Everyone is a developer only pays off if there is a layer underneath that absorbs the platform-grade work and leaves the app-grade work to firms.
That is the layer Giraffe is building. And the loop on top of it is what makes the pattern compound:
- A firm needs a custom utility - a feasibility flow, a schedule generator, a structural-cost estimator, a development-control checker.
- A senior architect or designer describes what the firm wants, in natural language, against the firm's libraries, materials, and conventions.
- AI writes the code that becomes the tool.
- The tool then runs inside Giraffe against the firm's models - without paying AI tokens on every invocation.
- The next project sharpens the tool. Edge cases get encoded. The senior's judgement accumulates inside the system.
- If public, another firm can fork the tool and modify it for their own conventions.
This is what turns AI into reusable office capability rather than per-project consumption. The AI built the tool. The tool runs on its own. The senior's expertise compounds inside the firm instead of walking out at retirement.
It is also the loop that makes the door-schedule problem from the last post solvable. The prompt is no longer a one-off - it produces code that produces the schedule. The next project starts from that code. The firm owns it.
An Old Pattern, New Economics
The historical analogy is worth stating plainly. Firms have been customising platforms to their internal standards for as long as AEC software has existed. CAD standards in the nineties. BIM templates and family libraries in the 2010s. Custom Revit add-ins. Dynamo and Grasshopper scripts. Every large practice has a folder of internal utilities that look broadly like what Giraffe is now formalising.
What is different is the speed at which a firm can produce one of these utilities and the breadth of problems it can plausibly cover. A custom Dynamo script in 2018 took a senior computational designer a week and required everyone using it to understand its quirks. A custom Giraffe flow in 2026, described in natural language by a senior architect and assembled with AI assistance against a documented API, can be in production in hours, owned by the firm, sanctioned as office capability.
The instinct is the same. The economics are different.
The Test I Have Not Yet Run
I have not yet put Giraffe through a project at depth. The orchestrator demos are compelling; the proof is whether a firm's twentieth custom flow is as easy to maintain as the first, and whether the trust the model assumes between vendor and firm holds at scale. Those are not questions a blog post can settle.
What I can say is this: of the directions on offer this year, the one where AI produces reusable office systems - not just one-off outputs, not just better individual demos - is the direction that compounds. Giraffe is the clearest illustration of it in production right now.