What to build as AI changes architectural work

At a glance TL;DR

• AI capability, software integration and adoption move at different speeds. A promising demonstration may be difficult to use on a live project, while a current limitation may soon disappear.
• Invest in foundations that remain useful as tools change: reliable project records, clear decisions, portable information and recoverable work.
• Test new workflows against the time and effort needed to reach an accepted result, including checking and coordination. Be cautious about bespoke systems built around temporary limitations.

A practice could spend months building around an AI limitation that soon disappears, or trying to make a promising demonstration work on a live project.

Both experiences are possible. They make the future architectural workstation a moving target: the combination of software, project information, interfaces, people and decisions through which work gets done. The opportunity is substantial. More of a professional’s intention may become executable without that person manually carrying it through every application. How quickly that becomes useful is a different question.

The opening discussion in Moonshots episode 283, published on 27 August 2026, offers a useful disagreement. Salim Ismail distinguishes technical possibility from institutional deployment, describing the organisational adaptation needed alongside stronger AI as co-scaling. Alex Wissner-Gross argues that accumulated abstraction layers, separating an intention from the systems executing it, can hold back change. His suggested remedy is vertical integration: bringing more of those layers together. Emad Mostaque counters that adoption also depends on models becoming good enough to use. The discussion’s impedance mismatch is a failure to connect growing intelligence supply with work that can effectively employ it.

Applied to architecture, these arguments suggest several interacting rates of change. Frontier capability determines what a model can attempt. Integration determines whether it can reach the right project information and act through useful tools. Practice adoption determines whether people can incorporate that work into delivery. Project and commercial arrangements influence whether saved effort becomes shorter programmes, better work or increased capacity.

Integration can make an existing model more useful by supplying context and reliable actions. A stronger model might make a complicated workaround unnecessary, operate an existing interface or help build its replacement. Practices experimenting with real work can reveal what software needs to support. These processes constrain and accelerate one another.

There is evidence of rapid progress, but it needs careful interpretation. METR’s task-horizon research, updated in May 2026, tracks increasing difficulty of tasks agents can complete, principally in software work. Its horizon measures the time a human expert would need for a task at a specified agent success rate. It does not measure how long an agent can be left working successfully. A result at 50 per cent success is also a long way from routine dependability. The observed trend cannot simply be transferred to coordinated architectural production.

Deployed behaviour supplies another view. Anthropic’s February 2026 study found that the longest Claude Code turns, at the 99.9th percentile, grew from under 25 minutes to over 45 minutes between October and January. The median remained around 45 seconds. A turn can end with a question or interruption, so duration does not establish successful completion. The authors suggest that changing usage and product improvements help explain the increase alongside model capability.

Nor does greater capability guarantee immediate productivity. METR’s February 2026 follow-up revisited its earlier finding that experienced developers took 19 per cent longer with early-2025 AI tools on their own projects. Later observations suggested improvement, but selection effects prevented a dependable estimate. This is not a verdict on current AI. It is a reason to measure completed work in context. Capability, error rates, running cost and sustained independent work are different quantities.

The ingredients continue changing. On 29 September, OpenAI added hosted-browser computer use to its Agents API. That expands a route for agents to act on websites; it does not establish dependable control of a Revit project. New access mechanisms and stronger models still need to meet the particular work.

Consider an illustrative request for information, or RFI, about a ceiling detail that conflicts with a structural beam. Producing a fluent reply is relatively easy to imagine. Completing the workflow means identifying the relevant drawing revisions, locating the structural advice, checking the site photograph and distinguishing an approved detail from an earlier option.

An assistant would need access to those sources and a way to show precisely what supports its proposed response. Missing or conflicting information should remain visible. The responsible professional still decides whether the response resolves the issue and who must approve it. After approval, an agent could potentially issue the response, update the register and link the resulting instruction to affected documentation. Each step requires access, verification and appropriate authority.

That is part of adoption latency: the work between demonstration and dependable use. Confidentiality, uncertain source versions and an inability to inspect or reverse changes are sound reasons to delay. Re-entering the same information into three registers, or requiring permission for every harmless lookup, may be avoidable friction. Removing it can require changes to the practice’s arrangements as much as to the model.

The abstraction-layer argument becomes concrete here. Project meaning is spread across email, documents, models, registers and conversations. A software provider that connects these layers may reduce hand-offs. A practice does not necessarily benefit from placing everything inside one vendor’s system, however. Useful boundaries preserve specialist responsibility; interoperable connections, exportable records and replaceable components preserve choices.

Autodesk’s current Revit Assistant documentation describes model queries and modifications, with Assistant updates delivered independently of yearly Revit releases. Its Revit MCP technical preview connects external AI clients to documented querying, reporting, export and navigation functions. MCP, the Model Context Protocol, provides a common way for AI clients to access tools and information. These are vendor-documented capabilities, not evidence that a whole coordination workflow is dependable.

A demonstration by AI Architecture Lab makes a bounded spatial workflow tangible. Existing rooms in Revit are displayed, furniture arrangements proposed and a person asked to choose. The sequence then shows furniture placement and 3D views for inspection. Its creator attributes the workflow to Claude connected through MCP; the particular implementation is unidentified. This is furnishing existing spaces, rather than establishing the design of an entire floor plan.

The edited footage leaves setup, elapsed working time, repeatability and checking effort unresolved. A practice would need to verify family dimensions, placement, circulation and affected documentation, and know how to recover from a failed change. The work could combine model reasoning with prepared automation. Its significance lies in shifting skilled attention towards choosing and checking while software carries out repeated operations inside an existing authoring tool.

A consultant revision illustrates the harder territory. Suppose a deeper beam affects ceiling levels and services clearance. A useful assistant must establish which revision supersedes which, locate the affected elements and views, and identify consequences across plans, sections and schedules. If permitted to prepare changes, it must respect coordinates, constraints and model relationships. Verification needs geometric checks and inspection of the affected documentation, followed by decisions about clashes and design consequences.

Reading the consultant’s note correctly is only the beginning. A convincing image or a valid software command does not establish a coordinated result. A June 2026 CAD research preprint illustrates one promising approach: combine AI planning with deterministic geometry tools and solver feedback. Its mechanical assembly prototypes are not evidence of architectural deployment, but they suggest why stronger language reasoning and dependable geometric execution may need to advance together.

The emerging proposition is intention, delegated action and review of results, alongside familiar authoring interfaces. Bounded spatial workflows are already being demonstrated. How reliably they generalise across projects and applications remains an open question; broad autonomous architectural coordination is still speculative.

Even successful task automation may barely change project throughput. In the hypothetical RFI, reducing response drafting from 30 minutes to five saves professional time. If the critical path still includes two days waiting for an engineer’s decision, the issue closes little sooner. If earlier checking instead discovers the conflict before a documentation deadline, it may prevent a much larger interruption. The location of the improvement matters.

Faster production can expose a shortage of reviewers, unresolved briefing or slow client decisions. It can also make checking more frequent and preparation more complete, reducing the work that reaches those bottlenecks. For a small practice, released hours can be significant even when the programme is unchanged. Under fixed fees they may improve capacity or margin; under hourly arrangements, the commercial consequences differ. Review, rework and operating costs belong in that calculation.

While the destination remains uncertain, clear revision state, accessible project records, the reasons behind important decisions and explicit definitions of acceptable work retain value across several futures. So do recoverable changes and connections that can accommodate another model or tool. These investments make work easier for people as well as agents.

Learning through a bounded implementation has value now. Choose a recurring workflow, test it against completed cases whose outcomes are known, and measure total time to an accepted result, including checking and correction. Record where assistance helps and where it creates work. Recheck those findings when the model or integration changes.

An elaborate system compensating for today’s model weakness is a more fragile investment. So is expensive automation tied to one changing interface. A modest implementation can still be worthwhile if it repays its cost quickly and can be replaced without rebuilding the practice around it.

Capability might continue outrunning integration, leaving impressive models constrained by fragmented project context. Integration might catch up quickly, making existing applications more readily delegable. Progress may also remain uneven, with geometric coordination lagging behind document work. The useful signals are fewer failed hand-offs, less human correction and more completed workflows, rather than a demonstration alone.

The workstation’s future form is uncertain. The shift worth preparing for is clearer: professionals spending less effort carrying information and instructions through software, and more attention deciding what should happen and whether it has happened correctly. A practice can prepare for that change without pretending to know its final interface.

Sources

1. Moonshots episode 283 transcript (third-party transcription) Retrieved 2 October 2026
2. METR: Time Horizon of Software Tasks AI Can Do Retrieved 2 October 2026
3. Anthropic: Measuring AI agent autonomy in practice Retrieved 2 October 2026
4. METR: Revisiting developer uplift after the early-2025 study Retrieved 2 October 2026
5. OpenAI API changelog: 29 September 2026 browser computer use Retrieved 3 October 2026
6. Autodesk Assistant in Revit Retrieved 3 October 2026
7. Autodesk: Revit MCP (Technical Preview) Retrieved 3 October 2026
8. Embodied CAD: Solver-Grounded LLM Agents for Parametric B-Rep Assembly Modeling Retrieved 2 October 2026
9. AI Architecture Lab: Revit x Jarvis furniture-layout demonstration Retrieved 2 October 2026