Today on The Anvil: a U.S.–Iran framework that both sides are describing differently in the same news cycle, a Garden Grove chemical tank that refuses to cool down with the DA now calling for whistleblowers, and Google quietly closing the Gemini CLI it spent a year crowdsourcing. Plus humanoids hitting 200 hours without a hardware failure, and a Spokane bonfire that became a five-acre reminder that fire season is already here.
Claude Code 2.1.147 ships three changes that reframe the tool from coding assistant to programmable infrastructure: a Workflow primitive for deterministic multi-agent orchestration (the agents now run as DAGs, not chats), Pinned Background Sessions that let an agent persist as a named service across invocations, and a refactored /code-review command. Combined with v2.1.150's per-category usage breakdown shipped Friday, the platform is now metering and orchestrating like a runtime, not a CLI.
Why it matters
The strategic move here: Anthropic is collapsing the GitHub Actions + LangChain + Copilot orchestration stack into one CLI for the 80% common case. For anyone running a small product team, the practical implication is that 'AI infrastructure' is becoming an ownership question — who owns the workflow YAML, who reviews subagent prompts, who pages when a Pinned Session degrades. The Cursor 3.3 parallel agents and Antigravity 2.0 Manager/Editor split (below) are pointing at the same architecture from different starting positions.
Figure AI's F.03 humanoid completed a continuous 200-hour autonomous run on a live warehouse conveyor sorting line, processing 249,560 packages at ~3 seconds each with zero hardware failures and no teleoperation. Control ran entirely on Figure's in-house Helix 02 neural net; the robots handled their own resets and energy management. Figure says its BotQ line is now producing one F.03 every 90 minutes.
Why it matters
200 hours uninterrupted on a real conveyor — not a demo, not a curated dataset — is the longest verified humanoid run in industry. It's the data point capital planners have been waiting for. The conventional 12–24 month payback math used for Locus or Amazon-style fixed automation now starts being defensible for humanoids in structured environments. Pair this with UBTECH's multi-robot Walker S1 deployment at ZEEKR's 5G factory this week and the question for 2027 capex is no longer 'are humanoids real?' but 'which tasks survive a human-labor cost-plus-benefits comparison.'
DeepSeek cut V4-Pro API prices 75% permanently on May 23, with input/output now at 0.025–6 yuan per million tokens (roughly $0.0035–$0.83). The company previously signaled prices would fall once Huawei's Ascend 950 supernodes shipped in volume in H2 2026 — implied but not confirmed. Separately, Andrej Karpathy joined Anthropic's pre-training team this week to build systems where Claude accelerates Claude's own training; and Google's Gemini 3.5 Flash API went GA at $1.50/M input tokens with 1M context.
Why it matters
Two distinct price-pressure stories converging: Chinese frontier-tier inference is now genuinely cheap (Huawei silicon making the geopolitical decoupling economically real), and Google is pricing Flash to undercut Claude/GPT for long-context workloads. For anyone running an evals budget against multiple model families, the cost-of-quality curve just shifted again. The Karpathy move is the more interesting research signal — the bet that AI-on-AI training acceleration is where Anthropic's edge gets compounded.
Microsoft Research released Fara1.5, an open-weight family of browser agents (4B / 9B / 27B), scoring 72% on the Online-Mind2Web benchmark against OpenAI Operator's 58.3% and Gemini 2.5's 57.3%. The 9B variant matches or beats proprietary frontier agents at a fraction of the parameter count. Permissive licensing for self-hosting.
Why it matters
Open-weight browser agents at parity with proprietary systems is meaningful for two reasons: it removes vendor lock-in for teams integrating agentic web automation into products, and it lets domain teams fine-tune for specific workflows (compliance, internal tooling, SOC2-restricted environments) without API-tier exposure. Combined with WebMCP shipping June 2, the browser-as-agent-surface is rapidly becoming a real platform — and Microsoft is positioning to own the open layer while Google focuses on the closed-source Antigravity stack.
Now on Day 3: the GKN methyl methacrylate tank is still heating at ~1°F/hour despite overnight cooling. Three escalations since yesterday's brief: OC DA Todd Spitzer opened a criminal investigation May 23 and is soliciting whistleblowers on the absent redundant cooling question; Governor Newsom declared a state of emergency, opening Costa Mesa Fairgrounds as a shelter; and at least six law firms have filed class actions on behalf of displaced residents and businesses. The ~40,000-resident evacuation zone remains in place. The Strawberry Festival is impacted.
Why it matters
The Spitzer whistleblower call before the tank is stabilized is the new signal — it means the DA is already working the theory that documentation of safety-system failures will emerge on its own. This is now a regulatory and legal accountability story running in parallel with the physical crisis, and it lands directly inside the OC Supervisor primary campaigns (June 2) already running on accountability themes. The broader question about industrial chemical storage inside OC's residential envelope is now formally on the political agenda.
Newport Beach launched five free, open-air retro trolleys May 23 covering 22 stops around the Balboa Peninsula, running weekends and holidays through Labor Day with real-time tracking via the TripShot app. Separately, Memorial Day observances across OC are leaning into the 250th-anniversary frame this year — Castaways Park is staging 1,776 American flags. Governor Newsom announced $540M in new transportation funding statewide May 23, with portions flowing to OC bridge and transit work.
Why it matters
Lighter context piece worth flagging if you're in Newport this weekend: the trolley network plus the Castaways display are the meaningful local-civic stories competing for attention against the Garden Grove crisis. The Newsom transportation tranche is too early to read for specific OC project allocations, but it lands as Dana Point Harbor and the Great Park expansion are both in active build phases.
Google's Gemini CLI — Apache 2.0, 100K+ GitHub stars, ~6,000 merged community PRs in its first year — sunsets for free and individual Pro/Ultra users on June 18, 2026. The replacement, Antigravity CLI (the same Antigravity 2.0 that topped ModelRift's OpenSCAD benchmark this week), is closed-source and enterprise-license-only, not yet at feature parity, with no 1:1 migration path and simultaneously dropping API access, model updates, and security patches for free users. Integration partners including Dynatrace, Figma, Shopify, and Stripe are caught mid-flight.
Why it matters
Apache 2.0 protects the code, not the cloud backend that makes it functional — the structural reality every team building on vendor-hosted 'open' AI tools should now price in explicitly. The Antigravity CLI closure makes this concrete in a way the abstract licensing argument never did: 6,000 community contributors just found out they were building customer-acquisition inventory. For anyone choosing between Antigravity CLI, Claude Code, Cursor, and self-hostable options, the backend-tier question is now the first evaluation criterion, not a footnote.
Two notable IDE-side shifts in the last seven days. Google's Antigravity 2.0 (I/O 2026) splits its UI into a Manager view (orchestration) and an Editor view (coding), with four parallel agents including a Chromium browser agent that visually verifies UI changes — but free-tier daily requests dropped from 250 to 20, a 92% cut since December. Cursor 3.3 (released May 7) adds parallel async subagents, automated PR-splitting that enforces review-sized PRs, and a native PR review tab inside the IDE. Meanwhile xAI's Grok Build entered the race May 14 with eight parallel sub-agents and local-first execution (70.8% SWE-Bench vs. Claude Code's 87.6%).
Why it matters
Three new architectural ideas landed in one cycle: (1) a browser agent that visually verifies CSS/UI changes — the closest anyone has come to closing the 'looks right vs. is right' gap that the Antigravity OpenSCAD benchmark exposed; (2) Manager/Editor as a first-class UX pattern for human-in-the-loop orchestration; (3) automated PR-splitting as a workflow constraint. For a design engineer, the browser-verification agent is the one worth a closer look — it's the first AI tooling that meaningfully addresses the design-output-quality verification gap rather than just generation speed.
Three concrete supply-chain AI updates this cycle. Blue Yonder shipped Pulse AI Dynamic Safety Stock — a microservice that autonomously adjusts per-SKU, per-location safety stock against real-time demand volatility and lead-time variability, replacing static 'set and forget' policies. ShipStation rolled out Carrier Intelligence with 11% average cost reduction and 0.4-day transit improvement in beta. Westwell demoed its E-Truck S2 and Q-Truck with the ReeWell dispatch platform at TOC Europe — already operational in eight countries from Pakistan to Mexico.
Why it matters
Three deployment stories with actual numbers, not announcements. Blue Yonder Pulse is the production version of the dynamic-buffer idea that's been theoretical in textbooks for a decade — and it sits naturally alongside the Conagra '95% no-touch planning' numbers from earlier this week. ShipStation's 11% cost reduction in beta is the kind of arbitrage on regional carrier pricing that Gartner specifically called out as one of the few near-term real wins last week. The pattern: where the data is clean and the optimization surface is well-defined, AI in supply chain is now delivering measurable, not aspirational, returns.
Chrome opened an origin trial at I/O 2026 for the HTML-in-Canvas API, which lets live DOM elements render directly into 2D canvas, WebGL textures, or WebGPU textures while preserving accessibility, find-in-page, translation, and extension support. Until now, web devs had to choose: semantic accessible DOM, or performant graphics. The trial runs alongside WebMCP (opening June 2) — both proposed by Google with Microsoft backing.
Why it matters
This is the first browser-platform change in years that unlocks new categories of application rather than refining existing ones: 3D product configurators with live, translatable, accessible text labels; design tools that don't need a custom text engine; WebXR with screen-reader-compatible content; agent-navigable 3D scenes. For anyone bridging physical product visualization and the web, this collapses the typical 'we built a custom text renderer for the canvas view' line item. Watch for the origin trial close date — that's when adoption commitments will start to firm up.
A rural Cheney bonfire escaped control Saturday afternoon in sustained 10–15 mph winds (gusts to 29), engulfed a building, and spread to roughly five acres by mid-afternoon. Nearly 40 units from Spokane, Whitman County, and Spokane FD 3 went defensive with helicopter water drops; a half-mile level-2 evacuation was issued by 5 p.m. The incident lands alongside an AccuWeather/NWS warning of an early-onset fire season across the Inland Northwest from below-normal 60-day precipitation. Coeur d'Alene begins a 13-acre Tubbs Hill phase-2 fuel mitigation Memorial Day.
Why it matters
Memorial Day weekend usually marks the start of casual-burn season; this year it's already in defensive mode. Worth noting alongside the Spokane renters'-right-to-cooling ordinance moving through council — the region is structurally hotter and drier than it was a decade ago, and the policy stack (rental cooling, fuel mitigation, prescribed burns) is starting to acknowledge it. Watch the Tubbs Hill closure schedule starting Monday.
Karen Cloninger, superintendent of Longview Public Schools and a former executive director with Spokane Public Schools, was arrested May 22 on charges of witness tampering, failure to report, and obstruction in connection with a sexual assault investigation at Mark Morris High School. Investigators allege district leadership had been aware of the allegations since January 29 but directed staff to handle the matter internally rather than report.
Why it matters
The Spokane Public Schools tie is what makes this a regional accountability story rather than a Longview story. SPS is already navigating a $2.5M budget deficit, a contested November/February levy decision, and 150–500 potential job cuts — and this surfaces uncomfortable questions about who in district leadership during Cloninger's SPS tenure was making similar judgment calls. Expect FOIA requests against SPS records in the next week.
Trump announced a 60-day MOU with Iran is 'largely negotiated': Hormuz reopens without tolls, the U.S. port blockade lifts, Iranian oil resumes, and nuclear talks defer 30–60 days. Tehran immediately called his account 'incomplete and inconsistent with reality' — insisting Hormuz stays under Iranian management and the Supreme Leader's ban on exporting the ~441 kg HEU stockpile (announced yesterday) stands, with Iran offering IAEA-supervised in-country dilution instead. Per Fars, the MOU includes mutual non-attack pledges and formally ends the Israel–Hezbollah front, though Israel reserves strike rights. ISW reads it as Iran negotiating from a winning-war posture: frontloading sanctions/Strait demands, slow-rolling the nuclear file.
Why it matters
The two structural gaps from yesterday's brief — Hormuz day-to-day control and whether HEU leaves the country — haven't moved; what's changed is that both sides are now publicly describing the same document in incompatible terms simultaneously, which is a new political constraint on the negotiation. Khamenei publicly blessing the MOU or the U.S. accepting dilution-in-place are still the tells to watch. Brent pricing in a partial breakthrough (~$103–$107) while the substance remains unresolved is the market bet worth monitoring.
New forensic depth on the Megalodon GitHub Actions worm from last week: Hudson Rock found 33% of affected accounts had a documented prior infostealer infection in their commercial dataset — and they estimate near-100% actual compromise of source accounts once unverified ones are included. The worm used the Shai Hulud framework leaked by TeamPCP; exposure data shows employees at Accenture, Adobe, and 24,000+ other companies have compromised GitHub credentials sitting in stealer logs now. Perplexity open-sourced Bumblebee this week, a read-only scanner targeting this exact threat surface.
Why it matters
The causal chain is now fully documented: commodity infostealer logs → leaked worm framework → CI/CD attack vector — no novel zero-day required. The TanStack SLSA-provenance defeat and the Nx Console VS Code extension compromise from last week hit the same upstream vectors (OIDC token extraction, marketplace trust). For any team running GitHub Actions, developer personal machines are the attack surface, and the new operational response is Bumblebee-class endpoint scanning, not just YAML hardening.
The 'largely negotiated' problem Trump and Tehran are publicly describing the same Iran MOU in incompatible terms — Hormuz reopens vs. Hormuz stays Iranian-managed, uranium leaves vs. uranium dilutes in place. Both sides need the optics; neither has conceded the substance.
Agentic IDEs are eating their own infrastructure Cursor 3.3 parallel agents, Claude Code 2.1.147 workflow primitives, Antigravity 2.0 with a browser-verification agent, and Grok Build's eight parallel sub-agents all shipped in roughly two weeks. The compression of orchestration into the IDE is now the competitive surface.
Open-source AI is a lease, not a deed Google closing Gemini CLI to free users after accepting 6,000 PRs is the most explicit version yet: Apache 2.0 protects code but not the cloud backend that makes it work. Anyone building on vendor-hosted 'open' AI tools should price in this risk.
Humanoids quietly cross the reliability threshold Figure's F.03 ran 200 hours autonomous, zero hardware failures, 249K packages on a live conveyor. UBTECH shipped multi-robot coordination at ZEEKR. The conversation is shifting from 'can they?' to 'what's the 24-month payback?'
Physical-world AI deployments are getting their failure stories too Starbucks killed NomadGo across 11,000 stores after nine months — workers had to recount every scan. The Garden Grove tank is binary leak-or-explode and no AI is helping. Reality remains harder than benchmarks.
What to Expect
2026-05-26—Memorial Day — OC observances (Castaways Park: 1,776 flags); Coeur d'Alene begins Tubbs Hill phase-2 fuel mitigation closures.