🛠️ Weekly Tech Brief

OpenAI launched Codex for Work on June 2, expanding Codex from a coding tool to a general office agent. Features include: one-click daily automations (morning brief, calendar review, email triage), "Skills" system for learning recurring workflows, and "Automations" for unprompted background tasks like issue triage and CI/CD monitoring. It integrates with GitHub, Slack, and docs.

PGC relevance: Codex for Work is a direct competitor/unbundled alternative to what Joe does. The "one-click daily automation" will feel familiar to anyone who gets morning briefs from Joe. The difference: Codex is cloud-only (your data goes through OpenAI), and costs per-user. Joe runs locally, free, with no data leaving the VPS.

Microsoft announced Scout at Build 2026 — an always-on AI agent embedded across Outlook, Teams, SharePoint, and the Microsoft Graph. Scout proposes, drafts, schedules, and follows up automatically. General availability: June 16, 2026. Simultaneously, a leaked internal planning document revealed phase 1's explicit goal was "make people addicted" to the agent. Microsoft issued a statement walking back the language but the damage was done.

The broader Build announcements include Web IQ (a search engine for AI agents), Fabric IQ (shared semantic space for enterprise data), and Foundry IQ (tying enterprise knowledge to the live web) — all launching June 16.

PGC relevance: Scout integrates deeply with PGC's likely Microsoft 365 stack. The "addiction by design" controversy is relevant as a governance question: whose data feeds the agent, and who controls the output? For a 34-person ESOP company, data sovereignty matters differently than for a faceless enterprise.

Google I/O 2026 introduced Managed Agents — a way to deploy and operate AI agents through the Gemini API without managing underlying infrastructure. Also announced: a more proactive Gemini app (agents that initiate, not just respond) and expanded agent creation workflows in AI Studio. The framing is "agentic Gemini" — agents as first-class services rather than chat interfaces.

PGC relevance: Google's managed agent approach makes it drastically easier to build AI agents without hosting them yourself. The trade-off is vendor lock-in and data residency. For PGC's MVP experiments, Google's approach is faster initial deployment, harder long-term exit. Joe represents the opposite trade-off: slower initial setup, zero exit cost.

O'Reilly's June 2026 Radar report identifies the central infrastructure question of the moment: "whether agents can transact and deploy without humans, and whether the platforms that host open source can stay reliable enough to keep that work going." Specific signals include: OpenAI and Qualcomm reportedly working on an agent-only phone (no apps), Kanwas (tool for sharing context across agents), Mike (open source local AI for legal work), and GitHub transitioning to usage-based Copilot billing.

PGC relevance: The "agent-only phone" concept — no apps, just an agent doing everything — makes Scout and Codex for Work look like early skirmishes in a longer war. For PGC, the question is whether to build agent infrastructure (Joe) that's independent of whatever platform wins.

OpenSpace announced (June 1) it has passed 1,000 data center projects processed through its Visual Intelligence platform. The system uses 360° cameras and smartphones to auto-map construction sites with time-stamped images, extracting metrics like percent complete, schedule progress, and productivity insights automatically. Integration with Procore and Autodesk is live.

PGC relevance: OpenSpace represents the production-ready version of what MVP 2 (Digital Counter with Togal.AI) was supposed to test — automated field intelligence. For PGC, OpenSpace is overkill for a 34-person office (it's designed for mega-projects), but its integration model (camera + AI -> Procore) is the direction construction tech is heading. Watch for a mid-market version.

BIM-Services.us announced (June 1) AI automation integration into its core modeling ecosystem, designed to eliminate pre-construction bottlenecks. The system automates clash detection, model coordination, and MEP/specialty system alignment that typically takes VDC teams weeks to complete manually.

PGC relevance: Glazing is one of the specialty systems that gets coordinated in the BIM process. If BIM-Services' AI reduces coordination from weeks to days, it directly impacts how PGC interacts with GCs and VDC teams. GCs will expect faster turnaround on model coordination and clash resolution. PGC that can match that speed wins.

ENR hosted a live event on June 4, 2026: "AI in Construction: From More Work to Better Work." The core thesis: VDC teams invest weeks in coordination (modeling, clash detection, alignment) but much of that work doesn't translate to the field. AI promises to close the gap between design coordination and field execution.

PGC relevance: The ENR framing is exactly the problem PGC faces — the gap between design intent and field installation. AI agents that bridge this gap (like Joe, or like OpenSpace, or like Scout) are the solution. The key insight from ENR: the bottleneck isn't AI capability, it's AI adoption at the field level.

Industry analysis (June 3) reports the U.S. smart window market entering a significant growth phase, driven by: rising energy-efficiency standards, smart infrastructure investment, and demand for occupant comfort. Technologies include electrochromic glass, dynamic glazing, and integrated building management systems.

PGC relevance: Smart windows are a premium product category that PGC could expand into. If the market is growing (and it is — multiple analyses confirm it), PGC has an opportunity to develop smart window installation capability before competitors do. The key question is training and certification — smart windows require different handling and integration than standard glazing.

Qubit Capital reports AI startup Series A rounds now average $51.9 million — about 30% above non-AI peers. Key verticals receiving funding: enterprise software, developer tools, construction tech, robotics, and healthcare. The analysis notes inflated valuations and questions about sustainable business models.

Why track this: Construction AI startups are raising at this level. It means the tools PGC might use in 1-2 years are being funded now. But it also raises the risk of an AI investment bubble — if funding contracts, some of these tools may not survive. PGC's strategy (test MVPs first, don't commit large contracts) is the right approach.

GitHub is transitioning Copilot to usage-based billing, moving from per-seat pricing to consumption-based models. The shift reflects the reality that not all developers use AI coding tools equally — heavy users were subsidized under per-seat pricing.

PGC relevance: This is a leading indicator for how AI tools across the industry will be priced. Expect construction AI tools (Togal.AI, OpenSpace, Scout) to eventually move toward usage-based pricing too. For PGC, that means: the tools that look expensive per-seat today may become cheaper as consumption models roll out — but variable cost also means variable budgeting.

IBM's 2026 AI analysis confirms the industry moving toward smaller multimodal reasoning systems tuned for specific verticals (legal, health, manufacturing). Mike (open-source local AI for legal work, from O'Reilly) is a leading example. Google's Gemini 2.0 and open-source models like Qwen 3.5 397B are all trending smaller, faster, and more specialized rather than larger and general.

PGC relevance: The trend toward smaller, specialized models is good for Joe. It means the models Joe runs (DeepSeek V4 Flash, Qwen 3.5) will get more capable without needing more compute. It also means industry-specific models for construction are coming. PGC could eventually run a glazing-specific model fine-tuned on PGC's own project data.

Best AI coding tool reviews for 2026 identify "agentic coding workflows" as the transformative development of the year. Unlike earlier AI coding assistants (suggested completions, code snippets), agentic workflows handle entire feature implementations: planning, coding, testing, and deployment as autonomous cycles. AI coding tools rated include Claude Code, Codex, Cursor, and GitHub Copilot.

PGC relevance: If PGC needs custom software (the glass calculator is the current example), agentic coding tools can build it faster. The PGC glass calculator was built with Joe's assistance — but a dedicated coding agent could take a spec and produce the full feature in a fraction of the time.

SpaceX secured a Google AI compute deal (following earlier Anthropic pact) ahead of its anticipated IPO. Marvell Technology joined the S&P 500 after the AI chip boom drove profitability. Both moves signal massive capital flows into AI infrastructure.

Why track this: The AI infrastructure buildout is so large that SpaceX — a rocket company — is now an AI compute customer. It means AI compute costs are unlikely to drop dramatically (demand keeps pulling supply). For PGC, this reinforces the local-first strategy: running Joe on Ollama avoids cloud compute costs that are structurally inflating, not deflating.

MIT Sloan Management Review's 2026 analysis flags familiar bubble markers: inflated startup valuations, hype-heavy media cycles, and massive infrastructure spending without clear ROI for most adopters. The piece doesn't predict a crash — it warns that enterprise AI investments need to show measurable returns, not just "we're using AI" signaling.

PGC relevance: This is the most strategically important article of the week for PGC. The "use AI because everyone is" logic is exactly what leads to overinvestment in tools that don't deliver. PGC's approach — test MVPs before committing, measure ROI, compare cloud vs. local — is bubble-resistant. The real risk is investing in an AI tool that doesn't survive a correction.

Kanwas is a new tool identified by O'Reilly Radar that enables sharing context across AI agents — essentially a collaboration layer for agent teams. Workgroups can use it to coordinate multiple agents working on the same project, sharing progress, decisions, and context without manual handoffs.

PGC relevance: This is the infrastructure for the scenario where PGC runs multiple agents: Joe (general), a construction takeoff agent, an estimating agent, a scheduling agent. Kanwas-like context sharing would let them coordinate without duplicating work or contradicting each other. For now, Joe is single-agent, but watching this category tells us when multi-agent coordination becomes practical for mid-market firms.