🛠️ Tech Trends Brief — June 1, 2026

15 Technologies for Pacific Glazing Corp

📅 June 1, 2026 🔬 15 Technologies ⚡ MVP Experiments 🏢 PGC Focused

⚡ Steve's Action Items

#ActionTool / TrendTime
1Try Ollama + nanochat — train + run a local LLM end-to-end, free, on your MacOllama + nanochat2 hrs
2Read the NSA MCP security advisory — agent tool protocol warnings from the fedsMCP Security15 min
3Watch Google Flow demo — AI filmmaking that could produce PGC safety videosGoogle/Veo Flow10 min
4Set up Claude Cowork for one task — let it autonomously file a spec or fill a formClaude Cowork1 hr
5Test agent-browser — automated data entry from web portals to your CRMagent-browser (Vercel)1 hr
6Check Bumblebee audit on PGC dev machine — supply chain scan, freeBumblebee10 min
7Calculate Copilot token budget — with new usage billing starting todayCopilot Token Billing20 min

🔍 Meta-Trends

Meta-Trend 1: The Agent Infrastructure Layer — Protocols, Memory, and Security Hit Production

Three things happened simultaneously this week: MCP/A2A protocols hit 97M+ downloads, the NSA issued a formal security advisory about agent protocol risks, and banks began testing agentic AI in production financial systems. Agent memory (OpenHuman, agentmemory) crossed 50K cumulative GitHub stars, making persistent private AI a real option. Supply chain auditing (Bumblebee, Perplexity AI) arrived just in time — MCP servers installed from Discord have no vetting. The infrastructure layer for agents is no longer theoretical — it's shipping with security gaps that need active management.

🟢 TRY — infrastructure stable enough for prototyping, manual audit required for production

Meta-Trend 2: The Physics of Cost — Token Markets, Local Inference, and the End of Flat-Rate AI

GitHub Copilot moved to usage-based billing today (June 1) — flat-rate is gone, tokens are metered. Dell reported 320x token consumption growth for reasoning agents. Small language models (Phi-4, Llama-3.2-3B, Qwen2.5-7B) running on-device via Ollama are now "good enough" for 80% of business tasks. nanochat (Andrej Karpathy, 55K+ ⭐) ships the full LLM training stack in one readable Python file — you could train a custom model for $100. Cloud API pricing is being replaced by token budgets. The economics favor local inference for "boring" business AI.

🟢 TRY — run the math on local vs. cloud for PGC's actual query volume

Meta-Trend 3: Computer Use Goes Commercial — the "Agent on Your Machine" Era

Three major players shipped "computer use" capabilities in overlapping timeframes: Claude Cowork (Anthropic) — sees your screen, uses mouse/keyboard, delivers finished files. OpenAI Operator — browser-based booking, form-filling, web automation. Google Gemini 3 — Computer Use tool in preview. Vercel agent-browser (7,000+ ⭐) — open-source browser automation CLI for AI agents. The "AI that uses your computer" is no longer a lab demo. Testing one this week costs nothing.

🟢 TRY — set up Claude Cowork or agent-browser on a sandbox computer today

📋 15 Technologies

1. Ollama + nanochat — Local LLM Infrastructure and Training in One Sweep Local AI
What it is
Ollama (170K+ ⭐) made local LLMs Docker-simple. Paired with nanochat (55K+ ⭐, Andrej Karpathy) — the full LLM training stack in one readable Python file. You can not only run but train a model on your own data, locally, for pocket change in compute.
What it's used for
Running LLMs on Mac/Windows/Linux without cloud dependency. nanochat teaches the full pipeline from data prep to sampling in ~1,000 lines of Python.
Trend tags
local AIOllamaLLM trainingopen sourceedge inference
MVP experiment
curl -fsSL https://ollama.com/install.sh | sh && ollama run phi4-mini:3.8b — then clone nanochat and train on glazing spec text: python train.py --data ./glazing-specs.txt. Time: 2 hours.
Why it matters for PGC
Local inference means zero per-query costs. A field tablet with Ollama + small model answers material questions, checks specs, summarizes RFIs without internet. nanochat proves you could fine-tune on PGC's own data.
🟢 TRY NOW
2. Bumblebee — Supply Chain Scanner for AI Tooling (Perplexity AI) AI Security
What it is
Read-only supply chain scanner (2.6K+ ⭐, v0.1.1). Checks npm, PyPI, Go, RubyGems, Composer, AND — critically — MCP servers, VS Code extensions, and browser extensions in one pass. Written in Go 1.25, zero non-standard dependencies.
What it's used for
Auditing AI agent supply chains. MCP servers, editor extensions, and browser extensions are the new attack surface that no existing scanner covers.
Trend tags
supply chainAI securityMCPdependenciesopen source
MVP experiment
go install github.com/perplexityai/bumblebee@latest && bumblebee scan. Time: 10 minutes. See every software risk on your dev machine.
Why it matters for PGC
As PGC adopts more AI tools (MCP servers, Claude Code skills, browser extensions), Bumblebee is the first line of defense. The NSA advisory this week specifically warned about MCP server security gaps.
🟢 TRY NOW
3. Copilot Token Billing — Usage-Based Pricing Starts Today (June 1, 2026) AI Economics
What it is
GitHub switched from flat-rate Copilot subscriptions to usage-based token billing today. Developers now pay per token consumed across inputs, outputs, and cached context. The old fallback to a cheaper model when quota is exhausted is gone.
What it's used for
Microsoft's move to monetize AI code generation at usage scale. Sets a precedent for every other AI coding tool.
Trend tags
token economicsGitHub Copilotusage billingdeveloper costs
MVP experiment
Review your GitHub org's Copilot usage for May. Calculate June cost under token model. Compare: what would same volume cost with local alternatives (Codeium, Continue + Ollama)?
Why it matters for PGC
If PGC developers use Copilot, this directly impacts monthly costs. It signals the industry-wide shift from flat-rate to pay-per-token — affecting every AI tool PGC might adopt.
🟡 WATCH + CALCULATE
4. Claude Cowork — Desktop Intelligence From Anthropic Computer Use / Agentic AI
What it is
Claude sees your screen and uses mouse/keyboard to automate software. Give it a goal ("file this change order") and Claude works on your computer, local files, and applications to return a finished deliverable.
What it's used for
Automating routine desktop workflows — filling forms, extracting data from legacy systems, routing documents. Different from chatbots: Claude does things on your machine.
Trend tags
computer usedesktop automationAnthropicagentic AI
MVP experiment
Pick one PGC task: "Open the current change order PDF, extract the glass quantities, and paste them into a new spreadsheet row." Let Claude Cowork do it start-to-finish. Time: 1 hour.
Why it matters for PGC
Legacy construction software often lacks APIs. Claude Cowork interacts with the UI just like a human — bridging the gap between PGC's old tools and new AI workflows.
🟢 TRY NOW
5. OpenAI Operator — Browser Agent for Web Automation Browser Automation
What it is
OpenAI's dedicated browser agent for booking, form-filling, and web task automation. Navigates websites autonomously, fills fields, clicks buttons, handles multi-step web workflows.
What it's used for
Supplier portal data extraction, subcontractor registration, permit applications, online bid submissions — any web-based form workflow.
Trend tags
browser automationOpenAIweb agentsOperator
MVP experiment
Go to a supplier portal PGC uses. Let Operator navigate to a specific order, check status, extract tracking number. Compare to manual time.
Why it matters for PGC
Subcontractor and supplier portals are notorious time sinks. Operator could automate 60%+ of web form filling that office staff handles daily.
🟡 WATCH — try when available
6. agent-browser (Vercel Labs) — Open-Source Browser Automation CLI Browser Automation
What it is
Vercel Labs browser automation CLI (7,000+ ⭐) purpose-built for AI agents. Fully open-source. Annotated screenshots, element targeting, clickable boundaries detected automatically.
What it's used for
AI agents that need to browse, extract data, fill forms — self-hostable, no per-query cost.
Trend tags
browser automationopen sourceVercelAI agentsCLI
MVP experiment
npx agent-browser screenshot --url https://supplier.portal.com --annotate. Then try navigating. Time: 30 minutes.
Why it matters for PGC
Open-source = no per-query cost, self-hosted, no data leaving PGC's network. Best option if PGC wants browser automation at scale.
🟢 TRY NOW
7. Veo 3 + Google Flow — AI Filmmaking for Anyone Generative Video
What it is
Google DeepMind's Veo 3 text-to-video model, now available through Google Flow — an AI filmmaking tool that creates cinematic clips from natural language. Google says this is the moment "AI video generation left the silent film era."
What it's used for
Creating short video clips from text: walkthrough animations, safety demonstrations, client progress updates, training videos.
Trend tags
AI videoVeo 3Google Flowgenerative mediamultimodal
MVP experiment
Go to labs.google.com/fx/tools/flow. Generate: "Time-lapse animation of curtain wall installation, glass units being lifted into place." Evaluate quality for client-facing use.
Why it matters for PGC
Progress videos for clients require cameras, editing, time. AI-generated video from text could produce weekly client updates in 5 minutes.
🟡 WATCH — quality improving fast, evaluate monthly
8. Computer Vision Breakthroughs — CVPR 2026: 4,090 Accepted Papers Computer Vision
What it is
CVPR 2026 accepted 4,090 papers out of 16,092 submissions. Key trends: monocular depth estimation (Depth Anything V2), 3D reconstruction from single images, and physical AI perception for robotics.
What it's used for
Automatic measurement from photos, construction progress tracking via drone imagery, defect detection on glass surfaces, safety zone monitoring.
Trend tags
computer visionCVPR 2026depth estimation3D reconstructionsite monitoring
MVP experiment
Install Depth Anything V2: pip install depth-anything-v2. Run on a single photo of a curtain wall installation. See if it estimates distances, detects gaps, measures alignment. Time: 1 hour.
Why it matters for PGC
Progress photos PGC already takes can be automatically analyzed: glass aligned to spec? Gaps in sealant? How much curtain wall installed this week? This is the visual layer of job site intelligence.
🟡 WATCH — ready for prototyping, evaluate against PGC site photos
9. Awesome AI Agent Papers 2026 — Agent Memory & Context Management Agent Research
What it is
A curated collection of 2026 agent research papers. Two standout contributions: "Beyond Dialogue Time" — temporal semantic memory organized by actual occurrence time. "Active Context Compression" — agent autonomously decides when to consolidate learnings and prune history, inspired by slime mold (Physarum polycephalum).
What it's used for
Building agents that remember what matters and forget what doesn't — without blowing their context window.
Trend tags
agent memorycontext compressionresearchtemporal memory
MVP experiment
Read "Beyond Dialogue Time" abstract on arXiv. If PGC's bid agent consolidated daily learning ("this client prefers 20-page proposals"), it stops needing re-prompting.
Why it matters for PGC
Memory is the unsolved problem in production AI agents. These papers show the path to "set it and forget it" agents that learn continuously.
📖 READ
10. Construction AI + Robotics — From Inspection to Automation Construction Tech
What it is
Three parallel tracks: MARIO project (Multi-Agent Robotic system for Inspection On-site — humanoid, quadruped, and aerial robots working together), Promise Robotics (AI-driven framing/assembly robots), and AI crew tooling claiming 8x productivity improvements through real-time site conditions and task sequencing.
What it's used for
Construction site inspection, automated framing/assembly, real-time crew coordination via AI.
Trend tags
construction roboticssite inspectionAI toolingphysical AI
MVP experiment
Watch MARIO project demo video. Map their multi-agent inspection approach to PGC site walk-throughs. Where could drone imagery + AI analysis replace a PM walking the site?
Why it matters for PGC
Construction gets robotic inspection. While PGC doesn't need humanoids for glazing, the AI analysis layer (progress tracking, defect detection from photos) applies immediately.
🟡 WATCH
11. Small Language Models (SLMs) — Good Enough for Business, Cheap Enough to Run Everywhere Local AI
What it is
Models under 7B parameters (Phi-4, Llama-3.2-3B, Qwen2.5-7B, Gemma-3) achieving GPT-3.5-to-GPT-4 level performance on business tasks — zero API cost when run locally. Gartner predicts 40% of enterprise workloads will incorporate industrial AI agents by end of 2026.
What it's used for
Document classification, spec extraction, material lookup, RFI routing — tasks where fast+cheap beats "best possible."
Trend tags
SLMsedge AIlocal inferenceon-devicecost efficient
MVP experiment
ollama run phi4-mini:3.8b — prompt: "Extract all glass types, thicknesses, and U-values from this spec paragraph." Compare to GPT-4o. Is it "good enough"? Time: 30 min.
Why it matters for PGC
80% of PGC's AI needs (spec reading, bid extraction, material lookup) handled by small models on a MacBook. No API costs, no data leaving the building, sub-second responses.
🟢 TRY NOW
12. NSA Advisory: MCP Security Gaps — Agent Protocol Security Matters AI Security
What it is
National Security Agency formal cybersecurity information sheet (May 2026, Ver. 1.0). Warning: "MCP's rapid proliferation has outpaced the development of its security model." Banks testing agentic AI in production — NSA worried about the attack surface.
What it's used for
Authoritative guidance on securing MCP-based agent deployments. Reference for evaluating AI tool supply chain risk.
Trend tags
NSAMCP securitycybersecurityagent protocolssupply chain
MVP experiment
Read the 12-page NSA CSI at nsa.gov. Identify which recommendations apply to any MCP server PGC is using or considering. Report: 3 biggest gaps found.
Why it matters for PGC
If PGC adopts MCP-based agents (Claude Code, Copilot, custom), the NSA guidance is the baseline security standard. Ignoring it is a liability with construction insurance requirements.
📖 READ — mandatory if using any MCP-based tools
13. Awesome Claude Skills — 1,000+ Production-Ready Skills for Claude Code AI Agent Skills
What it is
Massive collection (55K+ ⭐) of production-ready Claude Code skills. From design (transitions.dev — 12 CSS animations as a Claude skill file) to data processing, code review, documentation. Also: andrej-karpathy-skills — 4 Karpathy coding principles as one CLAUDE.md drop-in.
What it's used for
Extending Claude Code without writing skills from scratch. Drop-in CLAUDE.md files for specialized behavior.
Trend tags
Claude Codeskillsagent customizationopen source
MVP experiment
curl -o .claude/skills/karpathy.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md. Ask Claude Code a coding question — see behavior change. Time: 15 min.
Why it matters for PGC
If PGC uses Claude Code, these skills make it exponentially more capable. The skills ecosystem is the "app store" of the agent era.
🟢 TRY NOW
14. Apple Vision Pro — Spatial Computing Still Evolving for Construction Spatial Computing
What it is
Apple Vision Pro continues hardware evolution — all-black components surfacing online (second-gen signals). Spatial computing tech (depth estimation, hand tracking, eye tracking, 3D reconstruction) maturing for industrial applications.
What it's used for
AR-assisted installation, BIM visualization on site, remote expert guidance via spatial overlay.
Trend tags
Apple Vision Prospatial computingAR3D reconstructionconstruction tech
MVP experiment
Download AR model viewer on a tablet. Load a 3D BIM model of a curtain wall. Walk around. Is this useful for installation planning? Time: 30 min.
Why it matters for PGC: Spatial computing lets a glazier see the BIM model overlaid on the actual building — mullion locations, glass panel drop-ins, finished facade. As headsets get cheaper (next-gen likely under $2K), practical for site use.
🟡 WATCH — next hardware cycle brings cost down
15. Agentic AI in Enterprise — Gartner: 40% of Workloads Will Use Industrial AI Agents by Year-End Enterprise AI
What it is
Gartner: 40% of enterprise operational workloads incorporating Industrial AI Agents by end of 2026. Platforms: Kore.ai, Glean, Moveworks, Aisera, Sierra, Decagon, Cognigy. AWS Q Developer: 200+ AWS APIs autonomously. SAP Joule Studio: no-code agents pulling live ERP data, automating approvals.
What it's used for
Enterprise-wide AI agent adoption — IT helpdesk to supply chain to field operations. Standardized deployment patterns forming in real time.
Trend tags
enterprise AIindustrial agentsGartnerERP agentsSAP Joule
MVP experiment
Pick one enterprise platform PGC uses (SAP, Salesforce, Microsoft 365, Procore). Search "[platform] AI agent [year]". Map one PGC process to their agent framework. Time: 1 hour.
Why it matters for PGC
Enterprise vendors embed agents into every platform PGC uses. Question isn't "should we use agents?" — it's "which agents will be embedded in existing tools, and do we trust them?"
🟡 WATCH — vendor agents arriving in tools PGC already pays for

📊 Summary Table

# Technology What it is PGC Applicability Verdict
1Ollama + nanochatLocal LLM infrastructure + training🟢 High — run + train models locallyTRY NOW
2BumblebeeAI supply chain scanner (MCP + deps)🟢 High — audit AI tool installsTRY NOW
3Copilot Token BillingUsage-based pricing starts June 1🟡 Med — monitor if PGC uses CopilotWATCH
4Claude CoworkDesktop-level AI agent automation🟢 High — bridge legacy software gapTRY NOW
5OpenAI OperatorBrowser automation agent🟡 Med — web form automationWATCH
6agent-browser (Vercel)Open-source browser CLI for agents🟢 High — self-hosted, no per-query costTRY NOW
7Veo 3 + Google FlowAI filmmaking from text prompts🟡 Med — client progress videosWATCH
8CVPR 2026 VisionDepth estimation, 3D from photos🟡 Med — site photo analysisWATCH
9Agent Memory ResearchBeyond Dialogue Time, context compression🟡 Med — future agent architectureREAD
10Construction AI/RoboticsDrone + robot inspection, AI crew tools🟡 Med — inspection automationWATCH
11Small Language ModelsPhi-4/Llama-3.2/Qwen2.5, local inference🟢 High — 80% of tasks, zero costTRY NOW
12NSA MCP Security AdvisoryGovernment warning on agent protocol risks🟢 High — mandatory read if using MCPREAD
13Awesome Claude Skills1,000+ drop-in skills for Claude Code🟢 High — extend Claude cheapTRY NOW
14Apple Vision Pro gen 2Spatial computing for AR site use🟡 Med — next-gen cost reductionWATCH
15Enterprise Agent PlatformsSAP Joule, AWS Q, Gartner 40% prediction🟡 Med — embedded in existing toolsWATCH

🏗️ PGC Action Plan

This Week (30 min - 2 hours each):

This Month:

Next Quarter: