Construction AI: the guardrails catch up - agent containment, cheaper coding, and a planning-system warning

The Royal Town Planning Institute has warned that the UK planning system is not equipped to handle the way AI, automation and advanced manufacturing are reshaping land use and employment. The core problem is predictive: while AI and advanced manufacturing raise the economic value of an area, automation is simultaneously cutting the number of conventional jobs - so local planning authorities can no longer reliably forecast future employment levels or the associated economic value they are meant to plan for. The guidance underpinning these judgements, the Economic Needs Assessment (ENA) that authorities use for land allocation, was last updated in 2019 and predates the current wave entirely. The RTPI is calling for better integration of plans and strategies, a government-led National Spatial Framework, and regional collaboration to support emerging industrial clusters such as AI, advanced manufacturing and data centres. Alongside the warning, the institute has published best-practice advice for planners working with AI.

For anyone in development, infrastructure or data-centre work, this is the policy story that actually touches the pipeline. Employment forecasts feed land allocation; land allocation shapes what gets consented and where; and data-centre demand - the physical backbone of the AI boom - is colliding with a framework that can't model it. It is also a useful counterweight to the vendor optimism that dominates most AI coverage: the constraint on AI's impact in the built environment increasingly isn't the technology, it's whether the planning and skills systems around it can keep pace.

This is the direct sequel to last week's point about MCP being asked to carry a governance load it wasn't built for. Anthropic has extended its Claude Managed Agents platform with two enterprise controls. Self-hosted sandboxes (in public beta) let the actual execution of an agent's tools run on infrastructure you control - your own environment or a managed provider such as Cloudflare, Daytona, Modal or Vercel - while Anthropic still handles orchestration, context and recovery. That gives you control over network policy, audit logging, runtime configuration and, crucially, data residency. MCP tunnels (in research preview) let agents reach private MCP servers - internal databases, APIs, ticketing systems, knowledge bases - without exposing them to the public internet: instead of opening inbound firewall rules, you run a lightweight gateway that makes an outbound encrypted connection.

The plain-English version: these are the features that let an agent run inside your security perimeter rather than behind a sandbox your security team takes six weeks to clear. For construction, where project data is commercially sensitive and increasingly subject to Building Safety Act record-keeping and golden-thread obligations, "where does the execution actually happen and who can see the data" is the question that decides whether an AI pilot ever reaches a live project. This is the infrastructure that makes a defensible answer possible.

Computer-using agents in Microsoft Copilot Studio reached general availability (announced 13 May, rolling across all commercial Power Platform geographies including Europe, with government clouds following in the second half of 2026). The capability is the one most relevant to construction's back office: agents that operate websites and desktop applications through the user interface itself - using vision and reasoning to navigate live screens and adapt when layouts or fields shift - to automate processes that previously relied on brittle scripts or manual workarounds because the underlying systems had no API. The GA build ships with OpenAI's computer-use model and Claude Sonnet 4.5, Azure Key Vault for credential storage, Microsoft Purview audit logging, and configurable human-in-the-loop review.

Construction runs on exactly the kind of legacy, API-less software this targets - older accounting and cost systems, supplier portals, certification and compliance websites, plant-hire and procurement tools. The honest caveat is that UI-driven automation is inherently more fragile than an API integration and needs monitoring, but the credential-vaulting, audit-logging and human-in-the-loop features are precisely the governance scaffolding that makes it viable for regulated, auditable work. Note the recurring pattern across this week's releases: the headline isn't the model, it's the controls shipped around it.

Step back and the week tells one story. The capability question - can AI agents do useful, multi-step work? - was largely answered at I/O. The releases since have all been about the harder, more boring questions: where does the agent run, who can see the data, what gets logged, and who signs off? Anthropic's sandboxes and tunnels, Copilot Studio's credential vaulting and audit logging, even the human-in-the-loop defaults - these are governance features dressed as product launches. The RTPI's warning is the same lesson from the policy side: the technology has outrun the framework around it, and the gap is now the constraint.

For construction specifically, that is good news, because the sector's blockers were never really about capability - they were about assurance, auditability and data sensitivity on regulated projects. The tooling is finally being built to answer those questions. The teams that move fastest over the next quarter will be the ones who treat "show me the audit trail and the data-residency story" as the first question, not an afterthought - and who keep a human reviewing anything an agent produces before it touches a live project, a price or a programme.