Construction AI: Google I/O tonight, and Anthropic finally publishes its agent playbook

The main Google I/O keynote runs tonight at the Shoreline Amphitheatre from 18:00 BST (developer keynote follows at 21:30 BST), with the conference continuing through Wednesday. Confirmed agenda topics include Gemini, Android 17, Chrome, Cloud and "more". Expected: formal unveiling of Gemini 3.5 Flash and Pro (leaked checkpoints have circulated for a week with credible benchmark results), VEO 4 / "Omni" video model territory clarification, deeper Workspace agent integration (leaked screenshots last week), and Android XR for smart glasses ahead of Samsung's Galaxy Glasses at Unpacked in July.

For construction technology leaders, the practical advice is unchanged from yesterday: hold non-urgent AI commitments until Wednesday morning. The shape of Google's offering for the next 12 months will be substantially clearer by then, and several of the leaked items would change the right answer on video, multimodal and Android-tablet site deployment choices.

Procore reported Q1 2026 revenue of $359m (up 16 per cent year-on-year) with a 17 per cent non-GAAP operating margin. CEO Ajei Gopal used the call to position agentic AI as embedded platform capability - "autonomous AI-driven software agents embedded in Procore's platform to execute construction workflow tasks and decision-making, leveraging proprietary construction-specific data." The company expects increased sales of its AI offerings by Q3 FY2026, and ships a wave of platform updates today including a Planning feature beta with resource-specific Gantt assignments, and a redesigned submittal experience with automated date calculations.

For UK contractors and consultants whose clients are on Procore, this changes the buying conversation. The platform AI features are now part of the existing licence rather than a future add-on. Third-party AI tools that integrate via the Procore API will need to position themselves on either differentiated capability or specific UK-regulatory fit.

Anthropic's official Claude Code documentation now formalises the recommended four-phase methodology: Explore (Claude reads files, searches patterns, builds understanding - no modifications); Plan (Claude produces a detailed plan, opened in your editor for direct edit before execution); Code (Claude implements against the plan, with tests run and failures fixed); Commit (Claude commits with a descriptive message and opens a PR). The argument - backed by Anthropic and the Claude Code team - is that phases 1 and 2 are the cheapest in tokens and the most valuable in outcome.

This is the part to take away even if your firm does not use Claude Code: the methodology is portable to any agentic coding tool, and it gives engineering and digital teams a shared vocabulary for what disciplined AI-pair-programming actually looks like. The temptation to skip Explore and Plan is exactly where AI-assisted code review tends to fail.

A clear sub-pattern from the past week's tooling releases: developers now routinely run five or ten agents in parallel, and standard terminal multiplexers are not designed for it. Herdr (released this week) is one example of the new category - a terminal multiplexer specifically for AI coding agents, with split panes, scrollback per agent, and session attach/detach for asynchronous handoff. It sits alongside Claude Code's Agent View, Cursor's Background Agents and Microsoft Agent 365 in what is becoming a layered supervision stack.

A GitHub repository maintained by Lucas Valbuena (and the EliFuzz/awesome-system-prompts mirror) collects the leaked system prompts and tool schemas of more than 28-32 production AI coding agents: Cursor, Windsurf, Claude Code, Devin, v0, Lovable, Replit, Manus, Augment Code and many others. The repo has crossed 137k stars and 34.2k forks. Crucially, the JSON tool schemas are included - not just the prompt text - so what each tool has been permitted to do is visible alongside how it is told to behave.

This is the most useful free reference now available for anyone designing their own AI agent. Patterns to study: guardrail framing, tool-use protocols, refusal templates, file-system permission scopes, web-fetch restrictions, and how leading vendors handle the "agent claims success but silently failed" failure mode (the topic of Tejas Kumar's IBM "Harnesses in AI" talk this week).