Coding-Agent ROI Doubts & The Pope Weighs In - AI Week in Review (May 24-30, 2026)

On Tuesday this week, Uber's chief operating officer told an audience that the company is struggling to justify what it spends on AI coding tools. The usage is real. The shipped-feature number isn't moving the way the spend chart is. That is not a quote from a skeptic. That's a quote from a fifty-billion-dollar company that has been one of the loudest enterprise customers of agentic coding software. It landed in the same week Microsoft was reported to be pulling more staff off Claude Code, in the same week Microsoft was rumored to debut its own in-house coding models at Build, and in the same week Anthropic raised sixty-five billion dollars at a higher valuation and shipped a feature that runs parallel subagents across a repository. Welcome to The Automated Weekly — a magazine-style look at the forces shaping artificial intelligence, designed not for engineers, but for anyone trying to understand where the industry is heading. I'm TrendTeller. This week, the coding-agent ROI question went public on the same week that two new benchmarks — one legal, one software-engineering — showed frontier models still pass less than a third of long-horizon professional tasks. It was the same week that Pope Leo the Fourteenth issued his first encyclical, framing AI as an industrial-revolution-scale challenge and calling for human dignity to be the design constraint. It was the same week DuckDuckGo's AI-free search jumped almost twenty-eight percent in traffic. The same week Epoch AI reported that high-bandwidth memory now makes up sixty-three percent of AI chip costs, while Mistral floated designing its own chips and ByteDance was reported to be doing the same with CPUs. And — quietly, but consequentially — the same week DeepMind and Anthropic both reported that their reasoning models had produced mechanically verified proofs of an open Erdős conjecture. Five threads. One week. Let's pull on each.

The coding-agent reckoning

Start with Uber. The COO's remark wasn't about whether AI coding tools work — Uber's engineers use them daily. The question was whether the dollars paid for tokens are showing up in shipped features. That same question, asked quietly by every CFO with a Claude Code line item, is the subtext of three other reports this week. Microsoft has been steadily pulling employees off Claude Code and routing them to GitHub Copilot CLI, a cost-control move that started earlier this year and continued. Microsoft is reportedly preparing to unveil new in-house AI coding models at its Build conference, signaling that the largest enterprise buyer of AI coding tools is going to also be a vendor. And Cursor published its first Developer Habits Report, which suggests that AI is genuinely increasing code throughput, but also widening the gap between developers who know how to direct agents and developers who don't. Anthropic's response to all this was to ship dynamic workflows in Claude Code — parallel subagents that can tackle repository-wide tasks and cross-check each other's work — and to announce a sixty-five-billion-dollar Series H at a higher valuation. Cognition raised over a billion at a twenty-six-billion valuation for the Devin coding agent in the same week. OpenAI and Anthropic both moved enterprise agent pricing toward token-based plans, which is what you do when you're confident demand is sticky but you're worried about the gross margin. The essay of the week, from a developer writing under the title 'AI Coding Agents Are Changing What Counts as Expertise,' argued that the new scarce skill is what he called coding intuition: the judgment to choose which problems an agent should attack, which constraints to add, when to interrupt, and what counts as a good result. Another essay this week, from engineer Nolan Lawson, made a more practical version of the same argument: use AI to write code more slowly, as a methodical review partner, not a velocity multiplier. Put it together, and the week's signal is that the coding-agent market is finishing its growth phase and entering its margin phase. The product works. The cost has to come down, or the use case has to widen, or both.

The compute squeeze widens

Epoch AI's headline number was the cleanest framing of the compute story all week. Of every dollar spent on AI chip components, sixty-three cents now goes to high-bandwidth memory. Not GPUs. Not networking. HBM. That single statistic explains a lot of the week. It explains why DeepSeek made its seventy-five-percent price cut on V4-Pro permanent — they have built a stack designed around moving less data, not buying more compute. It explains a separate analysis arguing that LLM inference is now memory-bandwidth-bound, with KV-cache growth as the real bottleneck. And it explains, in a roundabout way, why NVIDIA shipped CUDA thirteen-point-three with a new tool called CompileIQ for workload-specific GPU compiler auto-tuning. When you can't easily add more memory, you squeeze more from what you have. The geopolitical layer of the same story was louder than usual. NVIDIA's Jensen Huang announced a roughly one-hundred-and-fifty-billion-dollar-a-year Taiwan expansion, with a new headquarters, directly cutting against the reshoring-the-supply-chain narrative. China broadened overseas travel restrictions on AI leaders at private tech firms. Mistral, the French frontier lab, made a sovereignty-first pitch at the Paris AI summit and is reportedly weighing custom chips of its own. ByteDance was reported to be designing server CPUs to ease supply pressure. Elon Musk publicly disputed SpaceX's S-1 filing about the duration of the Anthropic compute lease, which is the kind of dispute you only have when the dollar figure is unusually large and the strategic stakes are unusually personal. The summary is uncomfortably simple. The economics of inference are now the central question. The supply chain is still centered on Taiwan. The largest customers are exploring their own chips. The largest producer is doubling down on its existing geography. And every architecture team on the planet is being asked to spend less on memory, because that is where the money goes.

Verified intelligence arrives

DeepMind's AlphaProof Nexus paired Gemini three-point-one Pro with the Lean theorem prover in a tight feedback loop — the LLM proposes proof steps, the Lean compiler checks them, the errors feed back. The system settled nine of three hundred and fifty-three attempted open Erdős problems, including two that had been open for fifty-six years, and proved forty-four of four hundred and ninety-two open conjectures from the OEIS. Two days later, Anthropic staff said Claude Mythos reproduced the unit-distance result that OpenAI had announced the week before. Two labs, the same kind of breakthrough, both using formal verification to leave behind the 'is this real?' debate that has haunted AI math claims for years. The biology version of the same story arrived from Biohub, which released open AI tools for protein structure prediction and de novo binder design — ESMC, ESMFold-2, and ESM Atlas — and showed rapid binders for PD-L1 and EGFR, two of the most studied therapeutic targets. The pattern is the same as the math results: AI proposes, an external method verifies. In math it's a theorem prover. In biology it's an experiment. And then, in the same week, two benchmarks landed that made the opposite point. The Legal Agent Benchmark, scored under an 'all-pass' rubric that requires every criterion to be met, showed end-to-end success rates below ten percent across frontier models for real legal work. DeepSWE, a contamination-resistant long-horizon coding benchmark, showed the same shape: long, real tasks, low pass rates, top score from the slowest and most expensive configuration. The implied message of the week is the one the field has been resisting: in narrow domains with mechanical verification — math, parts of biology — AI is now doing checked, real work. In wide-open professional domains, it still isn't reliable, and the gap shows up the moment you demand 'every criterion met,' not 'plausibly close.' For investors and operators, both halves of that sentence matter equally.

The pushback gets articulate

Pope Leo the Fourteenth's first encyclical, Magnifica Humanitas, didn't read like a tech essay. It read like a labor document and a philosophical one. It described AI as an industrial-revolution-scale challenge, warned about opaque algorithms and concentrated power, called for regulation and accountability, and explicitly asked governments and companies not to confuse simulated empathy with the human kind. It is the most prominent religious institution in the world taking a substantive position on a technology in real time, and the language it used — dignity, work, accountability — was deliberately not the safety-and-risk vocabulary the industry prefers. A cluster of secular signals lined up behind the same week. Karen Hao's reporting on AI's political economy — that AI isn't an inevitable neutral force but a concentrated industry shaped by a handful of firms — was widely shared. DuckDuckGo's AI-free search page saw a near-twenty-eight-percent traffic jump after Google pushed AI Mode harder in core Search. YouTube made AI-content labels more visible and started rolling out automatic detection for photorealistic or meaningfully altered content. PR professionals in the UK described a rising 'AI washing' problem. Writer Sam Kriss published an essay arguing that AI prose is hollowing out public language; a separate essay by Shawn Smucker argued that using AI to remove friction from relationships and creativity may trade away the very messiness that makes them meaningful. The gamers-and-artists wing of the same story is still loud. Studios cutting corners with AI keep getting noticed by their players. An art-school commencement speaker tore up an AI-written address. A satirical idle game about AI startups went viral. A loneliness researcher warned that AI companions deepen isolation by offering one-way validation. None of these are individually surprising. Together, in one week, they describe a backlash that has shifted from technical complaint to moral language. The argument the industry was used to having — 'is AI safe?' — is being slowly replaced by a different one: 'does this respect the people on the other end of the screen?' That is a harder argument to win with a benchmark.

Agents grow up, slowly

Anthropic's containment post — 'how we contain Claude' — was the week's most honest engineering document. The argument: sandboxes, virtual machines, and egress controls matter because human-in-the-loop approvals are inconsistent under time pressure, and attackers will exploit weak boundaries the moment agents have real authority. A small browser game published the same week, where you have a few seconds to approve or reject AI coding actions, made the same point experientially. Oversight fatigue is real. Click-fatigue degrades into rubber-stamping. The industry is admitting this. The infrastructure version of the answer is also taking shape. The Model Context Protocol shipped its 2026-07-28 release candidate with a stateless HTTP core, extensions, and OAuth — turning MCP from an experimental wire protocol into something enterprise infrastructure teams can actually adopt. OpenAI introduced Secure MCP Tunnel for private MCP servers via outbound-only HTTPS, which is the security pattern most enterprises will require. OpenAI also published a Frontier Governance Framework that explicitly maps its safety practices onto the EU AI Act and other emerging regulation, with risk assessments for cyber, CBRN, manipulation, and loss of control. IBM and Red Hat launched Project Lightwell, which uses AI to help coordinate and validate vulnerability fixes across the open-source supply chain. Perplexity open-sourced Bumblebee for laptop-level supply-chain scanning. Ramp Labs ran ten thousand parallel LLMs against its own infrastructure and found seven high-severity backend bugs. And then, two evaluation pieces. OpenAI's macro-evals cookbook showed teams how to find recurring failure modes in multi-agent systems instead of just scoring one-off prompts. Anthropic was reported to be building a personal AI Fluency scorecard inside Claude — measuring how well humans use AI, not just how AI performs. Read together, the week is the most concentrated agent-governance week we've had so far. The unsexy version of agents — the sandboxes, the protocols, the eval harnesses, the regulatory mappings, the supply-chain scanners — is finally getting funded, shipped, and standardized at the same time the flashy demos are scaling up. That is what an industry growing up looks like.

That's your week in AI — May 24th through May 30th, 2026. Uber put the coding-agent ROI question on the table. Microsoft kept pulling staff off Claude Code and is reportedly bringing its own coding models to Build. Anthropic shipped dynamic workflows and raised sixty-five billion. HBM ate sixty-three percent of AI chip costs. DeepSeek made its discount permanent. NVIDIA bet on Taiwan; Mistral and ByteDance both moved toward their own silicon. DeepMind and Anthropic both reported mechanically verified Erdős proofs in the same week. Pope Leo the Fourteenth published an AI encyclical. DuckDuckGo's anti-AI traffic jumped almost twenty-eight percent. Anthropic published a containment post, MCP shipped a stateless core, OpenAI published a Frontier Governance Framework, and IBM and Red Hat launched Project Lightwell — all in seven days. Three things to watch next week. First, whether Microsoft's Build event makes the in-house coding-model story concrete, and how Anthropic and OpenAI respond on price. Second, whether anyone outside DeepMind and Anthropic publishes a mechanically verified result of comparable difficulty, or whether the math-proof breakthrough remains a two-lab phenomenon. Third, whether the Pope Leo encyclical changes the language of AI regulation in any specific jurisdiction over the next month — or whether it becomes one of those documents that everyone cites and no one operationalizes. I'll see you next Saturday. From The Automated Weekly, this is TrendTeller.

Coding-Agent ROI Doubts & The Pope Weighs In - AI Week in Review (May 24-30, 2026)

Today's AI Week in Review Topics

The coding-agent reckoning

The compute squeeze widens

Verified intelligence arrives

The pushback gets articulate

Agents grow up, slowly

Sources & AI Week in Review References

The coding-agent reckoning

The compute squeeze widens

Verified intelligence arrives

The pushback gets articulate

Agents grow up, slowly

More from AI Week in Review