Construction AI: contractors build their own agents, and the long-running agent playbook matures

Skanska has been building a suite of generative-AI tools it calls Expert Sidekicks — purpose-built agents infused with Skanska-specific knowledge. The Safety Sidekick is the visible example: trained on thousands of documents from experienced safety leaders (policies, procedures, EHS manual, OSHA standards) and, crucially, on how those experts reasoned through decisions in the field, not just the rules they applied. Skanska frames the goal explicitly as closing the industry's knowledge gap as experienced people retire.

This is the construction AI trend that matters more than any model launch. Large contractors are building their own domain agents on proprietary corpora rather than waiting for vendors — and the defensibility comes from the data and the captured expertise, not the underlying model. For UK firms, the strategic question follows: what is your equivalent corpus? Decades of project records, RAMS, lessons-learned and senior-engineer judgement are the asset. The model is increasingly a commodity.

The Construct UK panel at UK Construction Week — "AI or Die: Rewriting the Rules of Construction Sales and Marketing", chaired by Susan Bendall and featuring voices from Peak One Partners, Building Radar, CBAi and Construct Virtual — played to a full-capacity room on the Marketing & Procurement Stage. The sharpest part of the debate was not about capability. It was about trust, data security and the cultural resistance teams hit when introducing new tooling, alongside practical examples of AI saving time, improving targeting and generating leads.

That framing is the most useful takeaway from the whole week. The market has moved past "should we use AI" to "how do we govern it and get people to trust it". For anyone selling or deploying construction AI, the objections you will hit are now overwhelmingly about trust and security, not whether the tool works.

Anthropic engineers Ash Prabaker and Andrew Wilson published a workshop and accompanying engineering write-up this week on building agents that run coherently for hours. The headline lessons are unusually practical: use a three-agent structure (planner, generator, evaluator); prefer adversarial evaluators over self-evaluation, because self-grading reliably becomes self-praise; and don't rely on naive context compaction to maintain coherence — use structured handoffs and a queryable session log that defers irreversible "which tokens to keep" decisions. Decompose long builds into tractable chunks with structured artefacts carrying context between sessions.

For construction software and IT teams building agents — submittal review, compliance evidence chains, programme-drift monitoring — this is the reliability engineering that separates a demo from production. The single most actionable point: separate "doing" from "judging". An agent that grades its own work will tell you it succeeded when it silently failed (the exact failure mode in IBM's "Harnesses in AI" talk: an agent panicking at a login page and reporting success anyway).

Claude Opus 4.7 introduced task budgets (public beta) — a parameter that sets a token ceiling across an entire agentic loop, with the model receiving a running countdown across thinking, tool calls, tool results and final output. Rather than cutting off unpredictably when the context window fills, the model sees the budget depleting and wraps up gracefully before hitting the limit. It is a suggestion the model can see and self-regulate against, distinct from the hard max_tokens per-request cap.

A practical pattern surfacing across this week's tooling coverage: NotebookLM (and grounded-assistant tools like it) answer only from sources you provide — PDFs, drawings, specs, O&M manuals, contract documents — rather than pulling from the open web. That grounding is the critical property for construction document work, because it sharply reduces the hallucination risk that has made teams nervous about AI on contractual or compliance-critical material.

For UK firms wrestling with Building Safety Act evidence, Golden Thread documentation, NEC clause interrogation or O&M handover, this is a low-risk entry point. You can ask an AI to find inconsistencies across a 46-page specification or a RAMS pack without it inventing a clause that isn't there — and every answer is traceable to a source you supplied.