Cold Open — HP takes its OpenAI bet company-wide

Sunday, June 28, 2026. We scanned 2,593 items off the wire overnight; three made the cut — and the lead is the quiet kind of news that ends up mattering most for builders: a Fortune 100 company moving agents out of the lab and into the org chart.

🎧 This is the print twin of today's Cold Open episode. Listen to today's episode.

The lead · HP takes its OpenAI bet from pilot to production

HP Inc. announced it will scale its OpenAI Frontier strategic partnership, following a run of successful pilots, into a company-wide deployment. The framing is plain in OpenAI's own words: this moves AI "across the organization in areas ranging from customer and partner-facing solutions and experiences, customer telemetry insights and reporting, employee productivity, and software development."

What makes this worth leading with isn't the partnership — it's the numbers HP let OpenAI publish. Since it started testing Frontier in February, the early signs were the kind builders recognize:

"It has been an amazing tool, and I am using it daily." — an HP engineer, quoted by OpenAI

The receipts: one engineer moved through 122 pull requests across 43 projects in a matter of weeks. A security team remediated several software bugs in a day — work they estimated could otherwise have taken up to a month. HP teams found immediate value in OpenAI's APIs, ChatGPT, and Codex "inside real everyday work."

Why it matters

This is the agentic-delivery playbook executed at enterprise scale, in the open — the same shape as last month's Endava case study, but with a hardware giant and hard throughput numbers attached.

The part to actually internalize is the framing. HP isn't describing a tool; it's describing an operating model. OpenAI pitches Frontier as a "connective layer," and the reason is the sentence every team hits the week after the demo works: "agents need to know which context to trust, which tools they can access, what actions they are allowed to take, and how their outputs will be evaluated over time." Access, context, governance, evaluation. That quartet — not any single model — is the real product once you go from one clever script to a portfolio of agents running in production. HP says it has more than 100,000 partners and over 80% of its business flowing through them; "self-service agents across store, partner, chat, and voice" is the scale that quartet has to survive.

The fine print

Two caveats before anyone re-plans a roadmap around it. First, these are HP-and-OpenAI's own numbers, on OpenAI's own blog, with no third-party audit — "122 PRs in weeks" is one engineer's anecdote, not a controlled study, and "a month of work in a day" is an estimate of work that didn't happen. Second, most of what's described is a promise about a portfolio of agents still being built, not something shipped to customers today. The pattern worth copying isn't the metric, it's the operating model: govern access, pin the context an agent can trust, and evaluate outputs continuously — because that's the part that doesn't show up in a launch-day anecdote.

Sources: openai.com

02 · The biggest cheating scandal in the Ivy League is an AI story

Roberto Serrano, a professor of economics at Brown for 34 years, says he found conclusive evidence that at least 50 students cheated on the March midterm of ECON 1170, his advanced mathematical-economics course — what EL PAÍS calls the biggest known academic-integrity scandal at Brown and across the Ivy League. His account of the aftermath is its own story: he says the university president's first response was silence, and the dean stayed quiet until he took the case to the Academic Code Committee, which acknowledged it as "a wake-up call." Separately, Princeton is ending a 133-year-old unproctored honor code — in place since 1893 — and bringing professors back into the exam room.

Why it matters. This is the same trust problem AI created in code review, in customer support, in every place where output now arrives faster than anyone can verify it. "A.I. has made deception easier and more remunerative than ever before," a recent Stanford graduate wrote in The New York Times, adding that he didn't know a single classmate who hadn't used it to get through an assignment. Two stories collide on today's wire: institutions racing to deploy agents (see the lead) and institutions racing to detect them are running the same unsolved race — can you trust an output you didn't watch get produced?

Sources: english.elpais.com · nytimes.com

03 · The "central bank of central banks" puts a number on the AI-bust risk

The Bank for International Settlements published its Annual Economic Report on June 28 and named the AI investment boom among the top risks to global financial stability — alongside record public debt and fragile bond markets. The mechanism it spells out is specific: should the hyperscalers slow or halt their aggressive pace of capex, "many borrowers across the supply chain could struggle to replace lost revenue and service their debt." The financing of the boom, it warns, looks increasingly reliant on debt and complex funding structures — the conditions for the kind of overinvestment that ends previous boom-and-bust cycles.

Why it matters. Read it against the lead. Two stories up, HP is wiring agents into its org chart on the strength of real productivity gains; here, the institution whose entire job is watching systemic risk is warning that the capex frenzy underwriting all of it could overshoot. Both can be true at once. For anyone betting a roadmap — or a career — on this curve, the useful takeaway isn't "bubble" or "not a bubble." It's that the BIS now lists AI next to sovereign debt. Plan for compute getting cheaper and for the possibility of an air pocket.

Sources: bis.org · cnbc.com

Also on the radar

• AI fatigue hits the front page. Two of Hacker News's top threads today weren't about a model — they were about wanting less of one: "We need tech news sources which exclude AI" and Better Images of AI, a project pushing back on the glowing-brain-and-white-robot cliché. The backlash is itself a signal worth tracking.
• Hack Your Summer. DJ Patil is backing a free, 4-week "high-velocity production sprint" for students shut out of a brutal internship market — build something real and public instead of waiting on an offer that isn't coming. (Simon Willison)
• Silicon without HBM. A Show HN white paper for Sophon's PFG-1 describes a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM at all — an attempt to route around the memory bottleneck that gates inference cost. Early and unproven, but the direction is the story. (Hacker News)

Trends in dev tools

What moved in the tools engineers actually ship with.

• Agentic delivery now comes with a stopwatch. The HP–OpenAI numbers — 122 PRs across 43 projects in weeks, security fixes compressed from a month to a day — are the latest entry in a growing genre: production case studies that quantify what agents do to delivery throughput. The pitch has moved from "look what it can do" to "look what it did, with receipts." (OpenAI)
• The new bottleneck is review, not writing. Jon Udell's post title says it plainly: "Doctor, it hurts when agents create unreviewable PRs." Don't do that. His fix, surfaced by Simon Willison, is to flip "human in the loop" into "agent in the loop" — keep your normal workflow and invite agents into it, rather than rubber-stamping a black box's output. As agents write more of the diff, the scarce resource becomes a human who can actually review it. (simonwillison.net)
• The terminal is still where agents live. A Show HN, Bash4LLM+, is a single-file, dependency-free Bash wrapper for LLM APIs — no Python, no Node, just curl and jq — to send prompts, stream output, and process files line by line from the shell. The CLI keeps winning as the lowest-common-denominator surface for wiring models into real work. (GitHub)
• Coding evals keep formalizing. A new arXiv paper, "Towards Evaluation of Implicit Software World Models in Coding LLMs," probes whether a model actually holds a coherent mental model of the codebase it's editing — exactly the kind of eval that matters once agents, not humans, own most of the diff.

Popular skills

Agent skills — portable folders of instructions a coding agent loads on demand — keep spreading well beyond their Claude Code origins.

• Anthropic's own framing: skills are how you "equip agents for the real world." Its engineering write-up describes Agent Skills as composable folders of instructions, scripts, and resources an agent loads only when a task needs them — progressive disclosure, so the context window isn't paying for expertise it isn't using.
• Your backend can bring its own skills. Supabase documents an agent-skills install — npx skills add supabase/agent-skills — that hands a coding agent Supabase's own development and security guidance on demand. (It surfaced inside this very newsroom's tooling today.)
• Skills are going cross-agent, and getting catalogued. Roundups now frame "must-have skills for Claude — and any coding agent," and curated directories are cataloguing them the way npm catalogs packages — a sign the SKILL.md format is becoming a portable, discoverable layer rather than a single-tool feature.

AI fun fact

In 1770, Wolfgang von Kempelen unveiled a chess-playing automaton — a turbaned wooden figure seated at a cabinet — that went on to beat, among others, Napoleon Bonaparte and Benjamin Franklin, and toured Europe and America for 84 years before a fire destroyed it in 1854. It was a hoax: a human chess master was folded into the cabinet, working the Turk's arm with levers and magnets. More than two centuries later, Amazon named its crowdwork platform after it and coined a phrase for exactly that trick — "artificial artificial intelligence": software that quietly outsources the hard part to a hidden human. On a day when one builder argues we should stop saying "human in the loop," it's worth remembering the loop has been in the cabinet the whole time. (Mechanical Turk, the history)

---

Tomorrow: another Cold Open before your coffee cools. Full stories, every day, at penguinalley.com.

Sources: OpenAI — HP Frontier · EL PAÍS — Brown · BIS Annual Economic Report · CNBC — BIS · Simon Willison — Jon Udell · Hack Your Summer · Bash4LLM+ · arXiv — coding world models · Anthropic — Agent Skills