Planning AI and agent workflows are the real story

A new Planning Inspectorate document now explicitly references revised guidance on the use of artificial intelligence, issued on 20 February 2026. That sounds dry. But, it matters because it shows AI is no longer sitting outside formal planning and examination processes, it is being named inside them.

For UK infrastructure teams, that is the kind of change that slowly turns into day-to-day process. If AI touches the documents, checks, summaries, or evidence handling in planning, then governance stops being abstract and becomes operational.

The government's Extract assistant for planning officers and councils is older news, but it remains one of the most important UK construction-adjacent AI references. It is still the clearest example of how AI can be aimed at a genuine bottleneck, namely turning messy planning documents into something quicker to process.

That is useful context because it shows the direction of travel. The value is not in flashy demos. It is in taking friction out of document-heavy public workflows that affect housing and delivery.

A UK-focused construction article says AI is being used on London projects to reduce rework and improve verification during delivery. That is the sort of story I trust more than grand claims, because it sits close to the actual pain on site.

If AI can help teams catch errors earlier, verify work faster, or reduce the back-and-forth on site records, then it is doing real work. Not theory. Not a roadmap slide.

A new report says inefficient construction training is costing the industry £11m a year. That is not an AI story on the surface, but it is one of the main reasons AI adoption stays patchy.

If your workforce is already spending too much time on poor training processes, then adding new tools without fixing capability will only create more noise. The industry still has a basic readiness problem.

The strongest wider AI story this week is not about a bigger model. It is about how OpenAI's internal teams are using harness engineering around Codex. The lesson is simple. The model is only part of the system.

What makes it useful is the surrounding structure, build speed, observability, skills, scripts, and enough context for the agent to do work without constant human babysitting. That is directly relevant to anyone trying to build AI into construction workflows.

The latest agent discussions keep circling the same idea: agent systems get better when they can learn from traces, refine their own skills, and use persistent memory well. Hermes, OpenClaw-style workflows, and trace-sharing efforts all point in that direction.

For construction teams, the lesson is not to copy the tools exactly. It is to notice the pattern. The winning systems are the ones that can remember, adapt, and work through messy real tasks without being reset every time.

Anthropic's latest move with Claude Mythos and Project Glasswing says a lot about where frontier AI is heading. The model is reportedly very strong on coding and cyber tasks, but it is not being released widely. Instead, it is being held back for a restricted group.

That matters because it signals a split. Some models are becoming public tools. Others are becoming controlled capabilities. For teams that build software, or rely on software supply chains, that is a meaningful shift in how access and risk will work.

Gemma 4 is getting serious traction, including on-device use and local deployment. That does not automatically solve construction problems, but it does make local, privacy-conscious AI more realistic for field and office workflows.

If you need something close to the work, and not always tied to a cloud subscription, this is worth watching. The local stack is getting far more practical.