The week the agents moved into the middle of the job

The change most likely to move a programme this month arrived without a press launch. Reported across Building, Inside Housing and the trade press in the week to 12 August 2026, the Building Safety Regulator has extended staged Gateway 2 applications to single-tower higher-risk buildings, those at least seven storeys or 18 metres. Until now the staged route was reserved for complex multi-tower schemes. A developer can now separate groundworks and foundations from the rest of the build, get the substructure approved on its own, and start piling while the superstructure design is still being finalised and checked. Interim chief executive Charlie Pugsley called the staged-applications guidance potentially "key" to how quickly building control approvals get processed.

The background makes it sharper. The 2023 Higher-Risk Buildings Procedures Regulations always allowed staged applications; the Regulator just wasn't taking them for single towers, applying a viability test that was never in the legislation and acted, as the lawyers have argued for a year, as a de facto bar on staged delivery. Read it alongside the approval rate hitting 82 per cent in the twelve weeks to 1 August and the two point the same way: the door is opening, and you're now handed a way through it in stages rather than one enormous pack.

The caveat's worth stating plainly. Each stage still needs its own Gateway 2 approval before that stage can start, so staging sequences the scrutiny rather than dodging it. Get the substructure evidence right and you're out of the ground weeks earlier; get sloppy and you've multiplied the number of times the Regulator says no.

Today's action: if you've a higher-risk scheme parked behind a full Gateway 2 submission, get your building control route to price a staged application on the groundworks this week.

On 3 August Arcadis announced a strategic stake in AI platform Nomic and a long-term commercial deal, off the back of a six-month trial across roughly 150 engineers in 12 countries running Nomic's agents on real delivery work. The agents handle drawing review against firm standards, code compliance checks, submittal review, RFI research and BIM coordination, plugged into Autodesk Forma and Bentley ProjectWise. Arcadis says 86 per cent of participating engineers reckoned it changed how they approach their work, with one workflow dropping days of document cataloguing to hours. Consultancy-reported figures from people trialling a tool their employer went on to buy, so directional rather than gospel.

The other half of the same bet: Endra, the Stockholm firm building AI that generates coordinated MEP designs, calculations and documentation, confirmed a $50m Series A led by Andreessen Horowitz and opened offices in New York, San Francisco and London on 31 July, with London pitched as the hub for Europe and the Middle East. Total raised sits around $75m with roughly 30 enterprise clients, vendor-reported. What it's chasing is the grind Revit leaves you with, the weeks of coordinating building-services layouts and producing the documentation, with the claim of a coordinated first pass in minutes. I'd treat "minutes" as a fast first draft a competent engineer still signs, not a design cycle.

The aside an editor might cut: when a firm the size of Arcadis takes a roadmap seat, the platform starts optimising for Arcadis, and the smaller consultancy buying the same tool later inherits a product bent to a much bigger firm's shape. The direction is right, the review-and-coordination layer is where AI earns its keep. The ownership question travels with it.

The procurement filter: before you buy any "AI inside your engineering" tool, ask what it does with your project data and whether you can take the data and your configuration out if you leave. If the answer's vague, that's your answer.

The first bit of daylight in a while. The RICS UK Construction Monitor for Q2 2026, published 6 August, put the headline workload net balance at minus 4 per cent, still negative for the fifth straight quarter but up from minus 12 in Q1. The forward look is where the mood shifts: twelve-month workload expectations jumped from +2 to +13 per cent, private residential moved from minus 2 to +6, and private non-residential from minus 4 to +11. Infrastructure led again at +34, up from +19, with energy the standout at +39 as the data centre and grid build-out pulls work along. Water sat at +23, roads and rail lagged, so the recovery is concentrated where the power and compute money is going.

But read what the respondents themselves flag as the drag, and it isn't demand. It's the Building Safety Regulator, Gateway approval times and planning delays, quarter after quarter. That's a different problem to a thin order book, and it happens to be the one AI genuinely helps with today: the evidence pack, the submission pulled together without three people and a fortnight, the programme that flags its own slippage. Hold this next to the staged Gateway 2 change above and the week reads as demand and process finally moving towards each other.

My advice is to resist reading one quarter as a boom. A net balance still in the red, with the Regulator named as the brake, means the smart move is getting your evidence clean now so you're ready when the tap opens.

The practical bit: if your order book is turning, the constraint is shifting to the approvals and admin around the work. Point your AI effort there first, not at the digging.

On 10 August Meta released Muse Glimmer, easy to scroll past and worth stopping on. It's a 30-billion-parameter model, open-weight under an Apache 2.0 licence, free to use and free to run yourself. The trick that matters is the footprint: under 20GB of memory, so it runs locally on an ordinary Mac or PC with a single consumer graphics card. No data centre, no monthly seat, no round trip to a lab's servers. And it isn't a toy; it's built for agent work, calling tools, reading and writing files, working through multi-step tasks. Meta reckons it beats similarly sized open models from Google and Alibaba on about half its benchmarks, a vendor table I'd take with the usual salt.

What that means on site is the bit an incumbent won't rush to tell you. For the first time you can point a capable agent at your own documents on your own laptop with nothing leaving the building. If you've ever paused over what your golden thread, commercial correspondence or client's data is doing in someone else's cloud, this is the answer to a question the sector's been quietly asking for two years. And after the AISI story above, the pairing does the work: local running doesn't make an agent safe, but it does mean the blast radius is your machine, your log, your off-switch.

It won't suit every job, a 30B model won't match the cloud flagships on the hardest work, and running your own kit brings its own faff. But the option now exists, and it's free.

Worth doing: ask your software vendor whether a local or private-hosting option is on their roadmap. The answer tells you how seriously they take your data sovereignty.

For a year the AI money in construction went into the office: estimating, scope review, drawing checks, the paperwork about the digging. This week it pointed at the digging. Bloomberg reported in early August 2026 that SoftBank is weighing a deal that could value Gravis Robotics, a Zurich firm spun out of ETH, at more than $500m. Gravis doesn't build machines; it retrofits the ones you already run, bolting on a kit that fuses LiDAR, cameras, GNSS positioning and hydraulic sensors so a standard excavator can trench, grade soil and manage stockpiles on its own. SoftBank would fold the stake into Roze, the AI-and-robotics vehicle it's assembling alongside its $5.4bn purchase of ABB's robotics arm.

Two caveats before anyone gets carried away. It's a rumour, unconfirmed, size and structure unsettled, so treat the number as a marker. And an autonomous excavator on a live UK site runs into banksmen, exclusion zones, temporary works and a CDM regime that doesn't much care how clever the hydraulics are. The comparison only goes so far, but think of it as cruise control: useful on the long straight repetitive stretch, still needing a human at the roundabout. Trenching a service run across an open plot is the motorway. Threading a grab between live services on a tight urban job is the roundabout.

Still, when an investor of SoftBank's size starts buying the autonomy layer for plant rather than the admin layer for the office, it's saying the labour shortage on the tools is now a bigger prize than the one at the desk.

The takeaway: same question you'd ask of any software, asked of a machine. Whose sensors are on your excavator, who holds the data they capture, and can you take the machine and the record to another supplier. Autonomy you can't unbolt is lock-in with tracks on.

An analysis dated 8 August puts London's data centre pipeline at roughly 760MW live across 99 sites, with more than 8GW sitting in the electricity connections queue. Read the two numbers together and the 2027 question stops being how much capacity developers want and becomes which schemes can secure a grid connection, get transformers on stretching lead times, and find the commissioning and building-services people to fit them out. Analysts reckon 35 to 40 per cent of announced global capacity is at risk of delay or cancellation, a warning flare rather than gospel. This is the Ofgem deposit story from a fortnight ago arriving in the order book, with programmes live across Docklands, Brent Cross, Hayes, Park Royal and the Slough corridor.

The procurement filter: if you're pricing data centre work for 2027, ask for the grid connection date and the transformer procurement status before you resource the bid. Those two dates tell you whether it's real.

Building on the 2026-07-28 MCP specification we covered in week 31, Autodesk is opening a path for certified third-party MCP servers to be called inside its own Assistant, on top of MCP already in Autodesk Construction Cloud and a Revit MCP in technical preview with Revit 2027. Read one way, welcome openness. Read as a builder, worth asking which way the pipe runs: can your agent reach your data through a standard you could point at a different tool next year, or only through a pipe the platform holds. The wiring behind the wall is usually what decides whether the lights work.

The discipline: when a vendor says its agent connects to your other tools, ask how. An open standard you can repoint is a different purchase from a private pipe.

Two standing developments keep arguing for challenger construction software from the incumbents' side. Autodesk moved multi-user subscriptions to named-user pricing and scrapped renewal discounts from 7 January 2026, ending the old floating-seat flexibility. And contractors keep reporting Procore renewals climbing 10 to 14 per cent a year; one firm, fifteen years a customer, called the last two multi-year deals "outrageous", with analysts reckoning a mid-market contractor pays $10,000 to $60,000 a year and uses maybe 30 per cent of it. That's the challenger thesis in one line: enterprise money for a suite you use a third of, on terms that tighten every renewal. I'll be plain about the other side too, a two-year-old platform doesn't match a Procore feature list, and anyone who says otherwise is selling. Most firms don't need the feature list. They need the ten things they actually do, done well, at a price that doesn't ambush them.

For your board pack: before the next incumbent renewal, price two of the broad AI-first alternatives on a like-for-like basis, published pricing, data-export terms, notice period. Even if you stay, you'll renew from a stronger position.