Construction AI: MCP becomes the AEC standard, Provision's preconstruction agent, and Claude's one-flag Ultra Code

Autodesk's announcements at DevCon 2026 close out the integration story the brief has been tracking through the past three weeks. The headline: Forma Assistant is being positioned as the AI orchestration layer across Autodesk's product surface, working with sub-agents like the now-GA Project Data agent and able to call certified third-party tools. The supporting plumbing is the part to study. Autodesk has released its first set of public MCPs - Revit MCP (tech preview, for accessing and inspecting model data), Fusion MCP (tech preview, for creating, modifying and inspecting 3D geometry), Fusion Data MCP (tech preview, for component properties and administration), Product Help MCP (tech preview, for product documentation) and Fusion Automation MCP (private beta, for running Fusion workflows from natural language). The Design and Make Marketplace has been updated so certified solutions can now be called directly by Forma Assistant - turning the marketplace into the discovery layer for agent-callable tooling.

The wider point is the one worth bookmarking. Procore relaunched its Common Data Environment on 1 June with Datagrid-powered agentic AI built in. Bluebeam Max went live on 19 May with Anthropic's Claude inside Revu via MCP. Autodesk is now exposing public MCPs across Forma, Fusion and Revit. Three of the dominant AEC software vendors are building on the same orchestration protocol within a six-week window - and the practical implication for UK construction is that MCP support is now the sensible procurement filter for any AI-adjacent purchase. If a vendor isn't shipping it (or planning to), they're betting against the way the integration layer is actually going. A year ago that would have been a defensible position; this week it isn't.

The honest caveat is the one MCP itself has been carrying since spring - it's an interface protocol that's being asked to carry a governance load it wasn't built for, which is why Anthropic's self-hosted sandboxes and MCP tunnels (covered 28 May) matter as much as the protocol itself. The procurement filter is "ships MCP and has a credible answer on credentials, audit logging, data residency and approval gates." Both halves matter; one without the other isn't enough.

Procurement filter for the next twelve months: Does the vendor ship MCP (or plan to)? Is there a credible answer on credentials, audit logs, data residency and human approval? Two yeses or walk away.

The most credible UK-relevant preconstruction play I came across this week is Provision. The co-founder story is the bit that earns it a closer look. Founded in 2022 by Luigi La Corte (civil engineer) and Brendan Ardagh (quantity surveyor), the firm has reviewed more than $100 billion of project value across 66,000+ construction documents, raised $7m to scale the platform, and now ships a focused suite for GC preconstruction - scope, risk and document Q&A. The headline product is the Scope Agent: it generates a complete scope-of-work package from drawings, specs, tables, notes and addenda in under sixty minutes, replacing 30-40 hours of manual review per bid, with every requirement linked back to its source document. The architecture choice that matters is "pulls from drawings, specs and addenda" - not just one document type - because it's the cross-document reconciliation that creates most of the manual workload in real bids.

The reason this lands now in the UK is structural. RICS reported in 2024 that 31% of UK quantity-surveying practices had unfilled vacancies. SME tender packages typically consume 12-15 working days, much of which is repetitive review work that doesn't make the human cost-estimating judgement any sharper. Compressing the review side by 90%+ doesn't replace the estimator - it gives them their time back to do the work the tool can't, which is the validation, the commercial reading, and the trade-by-trade judgement that wins or loses the bid. The wider category point is straightforward. Document and compliance work is the AI win that's already landed in UK construction; preconstruction estimating is the next one with structural drivers, and the labour-shortage case makes the business model defensible regardless of the AI hype cycle.

The caveat is the one that applies to any vendor estimating-time claim. The 30-40 hours figure is Provision's own benchmark and will look different on your drawing and document standards. Run a calibration project on a recent bid where you already have the manual baseline before committing - that's the only honest comparison.

Worth doing: If you bid more than ten packages a quarter, run a structured trial on a single recent bid with Provision (or a comparable tool like Togal.AI or Kreo). Use your own quality bar and your own drawing standards. The procurement decision belongs to your numbers, not the vendor's.

A small but practical UX update from Anthropic. Claude Code has shipped "Ultra Code" - a new setting in the effort menu (the ultracode flag) that combines xhigh reasoning effort with automatic Dynamic Workflows orchestration. In effect, it's a one-switch shortcut to the Opus 4.8 multi-agent feature I covered on 1 June: up to 16 concurrent subagents running in parallel within a single Claude Code session, capped at 1,000 agents per run, with Claude drafting the plan, dispatching the work, validating the outputs and reporting back. Available in the Claude Code CLI, Desktop and VS Code extension for Max, Team and Enterprise plans, plus via the Claude API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry. Early-access teams inside Anthropic are using it for codebase-wide bug hunts, profiler-guided optimisation audits, security audits, and large migrations and modernisation efforts.

The UX point is what makes this worth flagging. The Dynamic Workflows feature was already shipping; what changes with Ultra Code is the activation energy. Most engineering teams that could benefit from multi-subagent orchestration weren't using it because the setup overhead wasn't worth it for the work in front of them. A one-flag mode means it gets used on real internal-tooling work, which is where the time savings actually compound. For construction software teams running internal automation - integrations off your BIM stack, compliance-record processing pipelines, take-off helpers, project-record reconciliation - this is the right thing to try this month on a properly-scoped task.

A practical action: If you have a backlog item that's "ten things across our codebase that all need the same kind of update", try it under Ultra Code with a deliberately small first job. Time the difference against the manual baseline. That's your unit economics for everything that follows.