Today on The Anvil: agent protocols quietly standardizing across IDEs, supply-chain AI moving from pilot decks to production picks-per-hour, and Iran trading missile posture for managed Hormuz transit while great powers negotiate the terms. Plus a Pacific Northwest land-use fight that previews the next decade of data-center buildout.
JetBrains released ReSharper 2026.2 EAP with the Agent Client Protocol (ACP), an open spec for how external AI agents plug into IDEs, with Junie shipping as the first reference implementation. The pitch is explicit: ACP is to agents what Language Server Protocol was to language tooling — a way for any editor to talk to any agent without rewriting the integration each time.
Why it matters
If ACP gets adoption from VS Code, Cursor, and the JetBrains lineup, the agent layer commoditizes the way LSP commoditized intellisense — and the differentiation moves up the stack to skills, memory, and orchestration. Watch whether Anthropic, OpenAI, and the Cursor/Copilot teams sign on or quietly route around it; the next twelve months decide whether agents end up portable or proprietary.
Two parallel moves toward agent-fleet operations. Cursor shipped configurable cloud environments — multi-repo support, Dockerfile-based config with build secrets, 70% faster layer caching, agent-led environment setup, and environment-level version history and audit logs. Same day, GitHub put its Copilot Cloud Agent tasks REST API into public preview, letting teams programmatically kick off fanned-out refactors, repo setup, and automated releases for Business and Enterprise tiers.
Why it matters
This is the operational layer for parallelized agent work: reproducible environments on one side, programmatic task dispatch on the other. Together they make 'run 50 agents against 50 services overnight' a governed workflow rather than a Slack thread of screenshots. The audit-log piece is the tell — both vendors are now pitching agent fleets to procurement, not just to developers.
Anthropic published a detailed engineering postmortem on a six-week Claude Code quality regression between March and April. Three independent product-layer changes — a reasoning-effort downgrade, a caching bug that wiped reasoning history mid-session, and a system-prompt verbosity cap — each shipped through internal eval and dogfooding without tripping alarms because evals ran on fresh sessions and narrow workloads. All three fixed by April 20.
Why it matters
The honest part of the writeup is the failure mode: evals didn't catch it because real users have long, stale sessions with messy context, and internal testing didn't. For anyone running Claude Code in CI pipelines or unattended overnight batches, this is the canonical cautionary tale — silent model routing changes and prompt-layer tweaks can degrade output without any visible signal. The implicit ask is that consumers maintain their own regression suites against the production endpoint.
MiniMax open-sourced its MiniMax-01 series — including a 456B-parameter Text model and a vision-language counterpart — built on a Lightning Attention mechanism that replaces standard Transformer attention with linear-complexity variants. Context window: 4 million tokens (20–32x competitors), with claimed performance parity to frontier models and API pricing at $0.2/$1.1 per million tokens.
Why it matters
Long-context inference economics are the bottleneck for stateful agent systems — Fractile just raised $220M on that thesis this week. An openly licensed model that runs 4M tokens at a fraction of frontier pricing is a real shift in what self-hosted agentic infrastructure can look like, and Lightning Attention is worth reading on its own technical merits regardless of whether you ever run the weights.
UK chip startup Fractile closed a $220M Series B led by Accel, Factorial Funds, and Founders Fund to build inference-specialized silicon. The pitch is blunt: model capability has outrun the memory bandwidth needed to serve it, and the economics of frontier AI now hinge on inference latency and cost at long context (100M+ tokens), not training FLOPs.
Why it matters
This lands the same week MiniMax open-sourced a 4M-token model, Google Cloud shipped 15 TB/s object storage, and Sakana/NVIDIA released custom CUDA kernels squeezing 20%+ out of FFN sparsity. The whole stack is reorganizing around the inference economics problem — and the venture money is following. For anyone budgeting AI features against a P&L, the marginal cost of a long-context call is becoming the question that matters.
Bot Auto moved a 230-mile commercial load between Houston and Dallas with no driver, no safety observer, and no remote operator, hauling a live shipment booked through Ryan Transportation. The company is claiming it as the first fully humanless over-the-road truckload run on a live commercial freight lane in the U.S.
Why it matters
Driverless trucking has been 'six months away' for roughly six years. A real shipper-paid load on a high-density freight corridor — without the broker theatrics of a safety driver behind the wheel — is the kind of milestone the AV freight thesis has been waiting on. Verify the operational concept of operations (lane geometry, weather constraints, yard handoff details) before reading it as scale-ready, but this is the first credible 'it actually happened' data point.
Two warehouse-automation data points landed together. Symbotic disclosed 2.23 billion cases processed in 2025 across customer DCs with 200M+ autonomous-vehicle miles, 20% more miles per bot daily, and 84% of the fleet averaging 40+ miles/day. Separately, Brightpick's deployment at Superior Communications consolidated two warehouses into a 25,000 sq ft Tennessee facility running 50 Autopicker robots — 3,500 picks/hour and 1,000 order lines/hour with just 2–3 pickers per shift.
Why it matters
These aren't pilot numbers. They're throughput data from real customers across multiple years of operation, and they undercut the 'AI in supply chain is still vapor' framing that Cambridge and GEP/Darden have been justifiably leveling at the agentic-platform side of the market. The robotics side is shipping. The diagnostic gap — robotics works, agentic data layers don't yet — is now the cleanest way to read this market.
Perfetti Van Melle (Mentos, Chupa Chups) published two-year results from a forecasting overhaul with BearingPoint and Anaplan: sales forecast error dropped from 3.6% in 2023 to 0.3% in 2025, planning time fell 40%, and product write-offs dropped 63.5%. The shift wasn't a new model — it was scenario-modeling across functions and faster decision cycles on top of cleaned-up demand-sensing data.
Why it matters
Reads cleanly alongside this week's Ventagium piece and the Maersk/MIT forum — both arguing that data integration and process discipline, not algorithmic sophistication, are what move the forecast-error needle. Perfetti is the rare published case showing what 'do the boring data work first' looks like as a P&L outcome. Useful evidence for anyone defending the unsexy integration line item in next year's budget.
The managed-access doctrine documented yesterday has now produced its first visible beneficiary: China-flagged vessels are being waved through Hormuz under Iran's traffic-management protocols while Trump and Xi meet in Beijing with Iran explicitly on the agenda. ISW's May 13 special report adds texture — Iran has restored roughly 90% of underground missile storage and 70% of mobile launchers, largely consistent with the classified intel (~70% of prewar missile stocks intact, ~90% of underground facilities) that contradicts administration decimation claims. Iraq and Pakistan are reportedly complying with Iranian transit protocols, normalizing the chokepoint regime without a formal treaty. Netanyahu publicly claimed a secret UAE visit; the UAE foreign ministry flatly denied it. The operational legal detail: the Pentagon is reportedly preparing to rename the operation 'Operation Sledgehammer' to reset the 60-day War Powers Authorization clock if the ceasefire breaks.
Why it matters
The China passage under Iranian protocols while Washington negotiates over them is the new load-bearing fact — it transforms the managed-access doctrine from a unilateral Iranian claim into something a P5 member is operationally ratifying. The War Powers clock reset maneuver signals the administration is actively war-gaming a ceasefire collapse rather than treating the pause as durable. Brent and Singapore bunker fuel remain the cleanest early-warning indicators.
Grant County PUD is condemning farmland and homes along a 31-mile transmission corridor to serve the Quincy data-center cluster. Property owners say the utility chose a route affecting 34 homes over alternatives affecting fewer than 10 — citing speed to revenue — and is offering ~25% of fair market value as one-time payments for permanent easements.
Why it matters
This is the political fight that follows every hyperscaler power deal in the Inland Northwest, and it's now naming names. Sits directly next to CalEthos/TerraVolt locking up 55,000 MMBTU/day for Southeast Idaho behind-the-meter gas earlier this week. The regulatory dodge — 'physical-infrastructure-as-a-service' to keep costs off ratepayer balance sheets — doesn't solve the land-use problem, it just shifts where the friction shows up. Watch whether Washington legislators move on landowner-compensation reform before the next Avista or PUD route hearing.
Bonner General Health and Benewah Community Hospital CEOs detailed a compounding crisis: rising insurance prior-auth denials, Medicaid reimbursement cuts, and housing costs in resort-adjacent towns making it impossible to recruit staff — forcing reliance on expensive traveling nurses. 67% of rural Idaho hospitals ran negative margins in Q4 2025, up from 15% two years earlier. A $930M federal rural health transformation grant offers partial relief.
Why it matters
The dynamic isn't really about healthcare — it's about whether the region's growth model is internally consistent. Population inflow drives housing prices, which prices out the workers needed to staff the services the new population demands. Same pattern that's playing out in Coeur d'Alene's restaurant labor and Spokane Valley's school staffing. Worth tracking which hospitals consolidate or close first; that's the leading indicator on broader service-base decay.
Orange City Council voted 5-2 on May 12 to advance a 1% sales tax for the November ballot, plus explored a hotel bed tax bump (10% → 15%) and cannabis legalization, against a projected $20M structural deficit. Lands the same week as the OC landfill tipping-fee jump (53% July 1), CalFresh cuts hitting ~300K county residents June 1, and Huntington Beach's contested redirect of federal housing-grant dollars to a councilmember-linked nonprofit.
Why it matters
The OC fiscal squeeze is now structural across multiple cities — pandemic-era reserves are gone, federal program cuts are landing, waste and energy costs are stepping up, and the housing-affordability shortfall (only 17% of 2021–2024 builds were affordable, per Voice of OC) means the property-tax base isn't expanding. Newport Beach has its own RHNA pressure and a contested council election arriving in this exact context.
Agoda's engineering team published a technical retrospective on how a single component library (@drone-js/core) plus brand-specific theme packages (@drone-js/theme-*) powers multiple white-label travel partners. The split: shared component logic stays in core; brand differences live entirely in JSON-based design tokens (color, spacing, type, radius) that flow through CSS variables. New partner onboarding = new theme file, not new code.
Why it matters
Useful canonical reference for anyone running a design system at scale with multiple consumers. The pattern — tokens as the contract surface, components as the runtime — is also the pattern that AI design-to-code workflows are quietly standardizing on. Reads cleanly against this week's argument that shadcn/ui became the default design system precisely because tokens-as-substrate made it AI-tool-legible.
Treasury added Hangzhou-based MizarVision to the SDN list on May 10 after the firm published geolocated commercial-satellite imagery of U.S. B-2 and B-1 bombers operating over Iranian airspace during Operation Epic Fury. MizarVision responded by treating the sanction as marketing — launching a hiring drive framing the U.S. designation as validation of 'combat-grade engineering.' First-of-its-kind sanction of a private commercial OSINT provider.
Why it matters
This is the U.S. extending export-control logic into the commercial OSINT market itself — not just satellite operators, but the analytics layer on top. The precedent matters: every Western OSINT firm publishing on Russian troop movements, Chinese naval activity, or Israeli operations is now operating in a different risk environment. Watch whether the EU or UK follow with parallel actions, and whether Bellingcat-style nonprofits get any clarification on where the line sits.
Agent protocols start to look like LSP did in 2016 JetBrains' Agent Client Protocol, Cursor's environment governance layer, GitHub's Copilot task REST API, and Semaphore's agentic CI/CD all landed within 48 hours. The shared bet: agents become a portable layer that IDEs, CI systems, and design tools speak to over a standard, not a vendor SDK.
Supply chain AI is finally shipping numbers, not slideware Symbotic at 2.23B cases. Superior Communications hitting 3,500 picks/hour with two pickers. Perfetti Van Melle cutting forecast error from 3.6% to 0.3%. Bot Auto running Houston–Dallas driverless. The credible operators are now leading with throughput data instead of vision decks — and the cautionary research (Maersk/MIT, Ventagium) still says the bottleneck is data plumbing, not models.
Hormuz as a managed-access regime, not a closed strait Iran's framing of the strait has shifted from blockade threat to toll-and-protocol governance, and China's vessels are passing under the new rules while Trump meets Xi. ISW and Euronews both flag that Iran has restored ~70% of pre-war missile capacity — the leverage hasn't actually been degraded, just repackaged.
Inference economics, not model quality, is the next squeeze Fractile's $220M Series B on the explicit thesis that inference is the bottleneck. MiniMax-01 open-sourcing 4M-token Lightning Attention. Google Cloud Storage Rapid at 15 TB/s. Sakana/NVIDIA's TwELL kernels. The story is shifting from 'who has the best model' to 'who can serve long-context cheaply enough to matter.'
Data centers are becoming a land-use politics story Grant County PUD using eminent domain on 34 homes for a transmission line to Quincy's data center cluster — at 25% of fair market value — sits next to CalEthos locking up 55,000 MMBTU/day in Southeast Idaho. The Inland Northwest is becoming a national case study in who pays for hyperscaler power.
What to Expect
2026-05-19—Kootenai County GOP precinct election (141 candidates, three factions, 74 seats) and Lakeland/Kellogg school levy votes.
2026-05-19—BMW Group's second Additive Manufacturing conference opens in Munich with Ford, Hyundai, Nissan presenting.
2026-05-13—Newport Beach: Blom–Curry debate over the Civic Center Park police station siting.
2026-06-01—GitHub Copilot's credit-based pricing takes effect; federal CalFresh cuts hit ~300K Orange County residents.
2026-06-26—California's $1B electric medium/heavy-duty truck rebate program opens for applications.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
847
📖
Read in full
Every article opened, read, and evaluated
153
⭐
Published today
Ranked by importance and verified across sources
14
— The Anvil
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste