Today on The Anvil: the agentic-everything narrative meets its first real receipts. Starbucks scrapped its AI inventory tool after nine months, fresh research shows coding agents are shifting work rather than eliminating it, and Iran's new Strait Authority moved from institution to operational jurisdiction claim β map, coastal waters, and active toll talks with Oman β while diplomacy inches toward a number (12 vs. 20 years) that still doesn't bridge.
Cohere released Command A+ under Apache 2.0 β a 218B-parameter sparse mixture-of-experts model with 25B active parameters that runs on two H100s with W4A4 quantization (or a single Blackwell). The model targets agentic tasks, RAG, and 48-language multilingual workloads, with ΟΒ²-Bench Telecom jumping from 37% to 85% and Terminal-Bench Hard from 3% to 25% versus its predecessor. 128K context, multimodal text/image, tool-use optimized.
Why it matters
The open-weights frontier keeps closing on proprietary. A capable agent-tuned MoE running on two H100s makes self-hosted agent fleets a procurement-level decision rather than a research project β particularly for buyers nervous about Cursor Composer 2.5's China-origin Kimi K2.5 base. Combined with Cursor's pricing and Gemini 3.5 Flash's economics, the practical takeaway is that 2026 is the year you should be designing for model portability rather than vendor lock-in.
With the June 2 primary approaching, OC Board of Supervisors candidates in Districts 2, 4, and 5 staked out competing visions for handling homelessness β prevention and mental-health services versus enforcement and accountability β against a Point-in-Time count showing a 14% two-year decline but 6,300+ still unhoused. Separately, a Beverly Press follow-up confirms Huntington Beach's $10K/month penalties escalate to $50K/month starting June if the city fails to submit a compliant housing element by May 28.
Why it matters
Two through-lines for the county: who controls the policy frame on homelessness for the next four years (with Foley-Nguyen-Wagner board tensions still hot), and how California's RHNA enforcement actually bites when a charter city refuses to zone. Huntington Beach has now exhausted its appellate options including a denied SCOTUS petition, and the monthly escalation is the rare state lever with real teeth. Worth tracking how Newport Beach and Costa Mesa adjust their own housing elements in the wake.
Stack Overflow, citing Smartsheet research, reports enterprise automation intensity rose 55% year-over-year but developer workload hasn't lightened β it's denser. Cognitive load is shifting from writing code to reviewing, validating, and gathering context for code agents produce. Resolve AI launched a multi-agent platform targeting the downstream incident-response burden created by AI-generated code in production.
Why it matters
This lands in the same week Anthropic's 'Code with Claude' conference reported nearly half of attending developers already merging Claude-only PRs β the generation side of the ledger is moving fast, and the review/governance side is visibly lagging. The structural implication for teams running agentic engineering workflows: PR-review throughput, observability tooling, and incident-response automation are now on the critical path, not the periphery. Boris Cherny's 'agentic engineering' framing from the May 8 conference is maturing into a real operational problem, not just a positioning shift.
At Anthropic's Code with Claude developer conference in London (May 19β21), MIT Tech Review reports nearly half of attending developers had already shipped pull requests written entirely by Claude. Anthropic previewed 'dreaming' β Claude Code agents writing notes to themselves between tasks to improve learning β and broader agentic workflows that test and iterate without human intervention.
Why it matters
The half-of-attendees-merging-Claude-only-PRs data point is the practical confirmation of what the doubled rate limits and lifted Opus throttling from the SpaceX compute deal were designed to enable. 'Dreaming' is the more interesting new signal: cross-task memory persistence is the architectural piece most agent frameworks are missing, and it directly addresses the session-coherence problem that's been the ceiling on multi-file agentic work. Pair with Stack Overflow's decision-fatigue research and the review/governance gap becomes the defining product problem for the next cycle.
GitHub open-sourced Copilot for Eclipse under MIT license, exposing implementation details for completion, chat, next-edit suggestions, and agentic workflows. The release lands the same week Sherwood News documents Cursor overtaking Copilot in web traffic, with Microsoft execs including Jay Parikh internally warning that Cursor and Claude Code could displace both Copilot and the GitHub platform itself if adaptation stalls. Recent GitHub outages and a breach of 4,000 internal repos added urgency.
Why it matters
Open-sourcing a strategic IDE plugin is rarely a position of strength β it's usually a community-leverage play after market share has shifted. For builders evaluating AI coding tooling, the takeaway is that the IDE-native bet (Cursor, Claude Code) has decisively outperformed the platform-bundling bet (Copilot inside GitHub) on developer preference. Watch whether Microsoft restructures Copilot's IDE positioning at Build or just keeps shipping incremental changes.
Starbucks terminated its AI-powered automated inventory counting tool across North American stores on May 21, just nine months after a September 2025 deployment. The system used LIDAR and camera data to count beverages and milk but mislabeled and miscounted items frequently, prompting a reversion to manual counting. CEO Brian Niccol cited the need for standardization in supply chain processes.
Why it matters
This is the rare visible failure in the agentic supply chain narrative β a high-profile rollout by a sophisticated retailer killed publicly, not buried. It's the empirical companion to this week's GEP/Darden study showing 95% of supply chain AI initiatives fail to scale. The lesson isn't that computer vision can't count cartons; it's that the variability of real retail backrooms β lighting, occlusion, inconsistent SKU placement β punishes systems trained or tested in cleaner environments. Expect this case study to surface in every supply chain AI procurement deck for the next year.
Manhattan Associates launched Solution Design Studio, an AI tool that lets business users describe warehouse, transportation, and supply chain operations in natural language and converts those descriptions to live configuration across Manhattan Active applications. Internal testing reportedly configured most of ActiveWarehouse from external designs in minutes rather than months. Manhattan's services teams are already using it; broader customer rollout follows.
Why it matters
Manhattan is the gorilla in tier-1 WMS/TMS. If natural-language-to-config holds up in customer environments, it directly attacks the implementation-services moat that's defined the category for two decades β the same moat Blue Yonder, SAP, and Oracle depend on. Pair this with the GEP/Darden 95% pilot-failure finding and the pattern is clear: vendors are racing to make AI useful at the configuration layer rather than the decision layer, where most pilots are still stalling.
The Figma agent β announced alongside Google Stitch and the DESIGN.md open-sourcing on May 20 β entered closed beta on paid plans with library awareness, design-system grounding, and parallel-prompt execution. New this cycle: Anthropic/OpenAI integration paths (Claude and Codex) confirmed, and a practitioner critique from Balint Bogdan arguing the agent cleanly splits design work into automatable production tasks (bulk edits, system docs, content population) versus judgment work (strategic decisions, user-informed pattern-breaking) that remains stubbornly human.
Why it matters
The closed beta is the operational marker β paid customers can now ship with it, closing the loop on the bidirectional design-to-code thesis Figma has been building toward since March. The 'production vs. judgment' framing from Bogdan is the more durable signal: it maps directly onto the agentic design-system components-as-contracts argument, where semantic tokens and machine-readable constraints handle the automatable layer and human judgment handles the rest. For builders at the design-code seam, investing in design-system fidelity and tokens is now validated by both the tooling and the practitioner critique.
Two pieces this week converge on the same thesis: Saqueib argues that as AI becomes embedded plumbing, the hard frontend problems are async UX patterns, event-driven state modeling, cancellation/recovery flows, form-driven intent capture, and accessible streaming interfaces β not chat bubbles. UX Collective's Daniel Ruston frames it as the move from the 'Interface Era' to the 'System Era,' where designers choreograph agency between humans and probabilistic systems rather than optimizing layouts.
Why it matters
Useful frame for anyone building real products on top of LLMs. The shipping problems are not visual β they are state coherence under partial results, graceful cancellation, and making model uncertainty legible without making the UI feel broken. Practical implication for hiring and team design: a senior frontend engineer who's deep on Suspense, AbortController, and form-driven state is now more valuable than one who's deep on animation or visual systems.
Perpetua Resources secured a $2.9 billion U.S. Export-Import Bank loan β the largest under EXIM's 'Make More in America' initiative β to fund its Stibnite Gold project in central Idaho. The mine will produce antimony, a critical mineral with no current domestic U.S. source, used in munitions, semiconductors, and renewables. Operational target: 2029.
Why it matters
This is the federal government writing the biggest single check yet to break Chinese dominance of antimony (currently ~50%+ of U.S. supply). Combined with North Idaho's silver producers benefiting from the November 2025 USGS critical-minerals designation, the Inland Northwest is consolidating as the federal preferred site for the domestic minerals supply chain rebuild. Watch downstream effects on regional labor markets, permitting timelines, and infrastructure spend.
Spokane Public Schools disclosed a $2.5M budget deficit and said staff reductions are inevitable regardless of levy outcome β roughly 150 positions if a November replacement levy passes, 400β500 if it fails. The board is weighing November versus February election timing. Separately, the board unanimously approved renaming the Community School to Riverpoint High School, with a 2028 relocation to a new building at 501 N. Riverpoint Blvd. in the University District alongside SPS central administration.
Why it matters
Two threads from the same board meeting. The levy timing decision is consequential β a February vote would push staffing decisions into a worse position with later notification timelines. The Riverpoint relocation, paired with the district's $12.4M property consolidation, signals Spokane Public Schools is making a structural bet on the University District as its operational and educational center of gravity. For anyone following the city's growth-corridor logic from Plan Spokane 2046, this aligns with the transit-oriented hubs decision earlier this month.
Iran's Persian Gulf Strait Authority β stood up May 19 and first covered then β published a jurisdictional map this week extending claims into UAE and Omani coastal waters, and opened active toll talks with Muscat. Bahrain, Kuwait, Qatar, Saudi Arabia, the UAE, and Secretary Rubio publicly rejected the regime, with Rubio calling any tolling system 'unfeasible' for a deal. Pakistan's Asim Munir and Interior Minister Naqvi are in Tehran for a second mediation round; Iran is reportedly offering a 12-year nuclear suspension against a US demand of 20+ years. Supreme Leader Mojtaba Khamenei directed that the ~441 kg HEU stockpile must remain in-country β reaffirming the ban on export he formalized on Day 84.
Why it matters
The PGSA moved from institution to operational jurisdiction claim in under 72 hours β map with named coastal waters, active Oman toll negotiations, and a separate reported Bitcoin-backed maritime insurance scheme all land this cycle. The negotiation gap is now numerically explicit (12 vs. 20 years) but structurally incompatible on two red lines: uranium location and Hormuz governance. Iran International's reporting on 24x fertilizer prices and farm closures is the domestic pressure signal worth watching β it may force movement before Trump's strike threats do.
HelpNetSecurity's weekly roundup features Babel Street's Insights Investigator β the agentic OSINT platform that launched May 20 β now positioned alongside Trust3 AI's MCP Security product. The NSA's AI Security Center released formal security design considerations for MCP this week, calling out serialization vulnerabilities, trust boundary issues, and agent misuse risks that traditional cybersecurity tooling doesn't address.
Why it matters
The NSA sheet is the first formal threat-model document for MCP from a standards-level source β and it arrives just as ShadowBroker, OpenOSINT, and Babel Street's Investigator are all landing in production. The governance catch-up pattern running through this week (Gartner's agent-washing warning for supply chain, Microsoft's FIDES for coding agents, now NSA for OSINT/MCP) is consistent: capability moved faster than audit infrastructure, and the formalization wave is now arriving. For anyone running MCP servers in production, the NSA sheet is the document to pin.
On May 11, threat group TeamPCP compromised TanStack via GitHub Actions cache poisoning and OIDC token extraction, publishing 84 malicious versions across 42 @tanstack npm packages. The worm propagated to 160+ secondary packages including Mistral AI and UiPath. Notably, this is the first documented npm supply-chain attack to carry valid SLSA provenance attestation β undermining the integrity guarantee SLSA is supposed to provide.
Why it matters
Two things matter here. First, GitHub Actions cache and OIDC tokens are now confirmed attack surfaces at scale, which should prompt immediate workflow audits and credential rotation for any team running CI/CD on shared runners. Second β and this is the more durable point β SLSA provenance no longer means what builders thought it meant. If the build pipeline itself is compromised, attestation is just signed malware. The software supply chain security story for the rest of 2026 will be about hardening the build, not adding signatures.
The Agentic Hangover Is Here Starbucks killing its nine-month-old AI inventory tool, Stack Overflow's decision-fatigue research, and the GEP/Darden 95%-fail-to-scale finding are all landing in the same week. The narrative is shifting from 'agents replace work' to 'agents redistribute work' β usually toward review, judgment, and governance.
Hormuz as a Productized Chokepoint Iran's Persian Gulf Strait Authority moved from announcement to map this week β now claiming jurisdiction into UAE and Omani waters, holding toll talks with Muscat, and reportedly developing a Bitcoin-backed maritime insurance scheme. The strait is being converted from a military lever into a recurring-revenue product.
Cohere and Cursor Squeeze the Frontier on Price Cohere's open-source Command A+ runs on two H100s; Cursor's Composer 2.5 lands at $0.07/task at third place on the Coding Agent Index. The premium-frontier moat is shrinking to specific Opus/GPT-5.5 use cases β enough for builders to build cost-tiered routing as a default architecture.
Critical Minerals Reshoring Goes Inland Perpetua's $2.9B EXIM loan for the Stibnite antimony project is the largest under 'Make More in America' β and lands the same week silver producers in the Silver Valley benefit from USGS critical-minerals designation. The Inland Northwest is quietly becoming the federal government's preferred mineral supply chain bet.
Design's Center of Gravity Moves to System Choreography Figma's agent, Stack Overflow's review-burden data, the UX Collective's 'System Era' framing, and Designer Fund's 91% weekly-AI-use number all point the same direction: the work isn't pixels, it's partitioning agency between humans and probabilistic systems. Frontend skills that matter now are streaming UX, cancellation, and intent capture β not chrome.
What to Expect
2026-05-25—Tubbs Hill Phase 2 fuel-reduction project begins; trails closed in Coeur d'Alene.
2026-05-27—Newport City Council takes up FY 2027 budget, school deficit communications, and harbor recommendations.
2026-05-28—Huntington Beach housing-element compliance deadline before $50K/month penalties activate in June.