AI the News That's Fit to Prompt

Anthropic buried the most striking number of the day inside a product launch: 65% of its own product team's code is now written by an internal version of Claude Tag, the @-mentionable Slack teammate it just shipped to everyone else. That is the through-line of June 25, the conversation has moved past whether AI can do the work to how you hand it the work, who it answers to, and what it costs to run. From OpenAI bending metal to lower the price of inference, to a 51-page Google playbook on "agentic engineering," to data showing engineers are the hardest job to kill, today is about AI graduating from a thing you prompt into a system you operate.

The AI Teammate Clocks In

• Claude Tag makes the agent a coworker, not a chatbot. You delegate by typing @Claude in a Slack channel; it inherits the channel's context, uses connected tools under its own logged-in account (so every credential is auditable), and persists across days. Permissions are scoped per-channel: tag it in the legal channel and it can read contracts but literally cannot touch the codebase, because it cannot see it from there. It replaces the old "Claude in Slack" app with a 30-day migration window, in beta on Team and Enterprise plans.
• The 65% figure reframes the whole pitch. Anthropic's own product team already gets most of its code from this tool. That is not a demo, it is a disclosure about how the company building the model now builds software, and it sets an uncomfortable benchmark for every engineering org watching.

The Money and the Metal

• OpenAI's first custom chip is named Jalapeño. Built with Broadcom (a partnership first announced October 2025, with OpenAI's own models assisting the design), it targets inference, the cost of running models, not training, which stays on Nvidia. OpenAI claims "significantly better performance-per-watt than current state-of-the-art alternatives," tuned for real-time coding models. It is still in testing with no ship date. The logic, in OpenAI's words: "OpenAI operates across the stack, each layer can be optimized around the same goal."
• Mistral OCR 4 attacks cost, not just accuracy. It returns structured output (bounding boxes, typed blocks, per-word confidence) across 170 languages at $4 per 1,000 pages. One engineer at Rogo reported matching leading agentic document parsers "at roughly 8x lower cost and 17x lower latency." The unglamorous plumbing of agents, reading messy documents, is where the next cost war is being fought.

Agentic Engineering Stops Being a Vibe

• Google's 51-page masterclass has a blunt thesis. Per Cole Medin's breakdown: the model is roughly 10% of your system, and the harness (context, tools, guardrails, verification, observability) is the other 90% and matters more. Evidence cited: an AI layer lifted a sub-top-30 model into the top 5 on Terminal Bench 2.0, and LangChain reportedly added 13.7 points, about the Sonnet-to-Opus gap. Vibe coding is recast as one end of a spectrum (fine for disposable work); agentic engineering, with specs, evals, and CI gates, is the other.
• Three creators converged on the same idea from different angles. Chase AI maps an "agentic OS" where 90% of the value lives under the hood in skills, loops, and a well-structured file system, and separately breaks loop engineering into four phases (trigger, execution, verification, state) where the success criteria decides whether a task even belongs in a loop. Nate B. Jones reframes it as a "loop of loops," recurring jobs that remember, notice change, and crucially stop before acting where your judgment is required. The shared insight: the unit of work is no longer the prompt, it is the durable, supervised loop.

Models, Tools, and Defense

• Google shipped a wave. Gemini Omni Flash generates and edits video from mixed image/audio/video/text inputs; computer-use landed in Gemini 3.5 Flash for cross-device agents; Gemma 4 12B is an encoder-free multimodal model that runs on a 16GB laptop; and experimental DiffusionGemma claims 4x faster text generation by producing blocks at once instead of token-by-token.
• OpenAI's Daybreak points its models at defense. Bundling frontier cyber models, Codex Security, and partnerships to help defenders find and fix vulnerabilities first, it is a deliberate counterweight to rising anxiety about the same models' offensive cyber capabilities.

Where AI Meets the Body and the Career

• Microsoft's Talos turns old genomes into new diagnoses. It re-interprets a patient's existing variant data against the latest knowledge each cycle. Across 4,735 undiagnosed individuals it produced 241 new diagnoses in 238 people (a 5.1% additional yield), every one later confirmed pathogenic by accredited labs, averaging just 32 days between new evidence appearing in a database and a patient getting an answer.
• Engineers are the hardest job to kill, so far. SignalFire found total big-tech hiring fell 25% versus 2019, but engineering roles fell only 11%, and engineers were 55% of all new hires in 2025 (up from 46%). "They're suddenly a lot more productive, and there's endless work for them to do," said research head Asher Bantock.
• Dylan Field's contrarian read on AI and design. "AI draws from the middle of the distribution, and to be differentiated you need to be at the tails." Figma's bet is that more activity on its canvas feeds a data flywheel model providers lack, and that "Code on the Canvas" collapses design and code into one continuum rather than rivals.

The Throughline

For two years the AI story was the model. Today, almost every story is about everything around the model. Anthropic's 65% code figure, Google's "the harness is 90%" thesis, Chase AI's agentic OS, and Nate Jones's loop of loops are four versions of the same claim: the model is now a commodity component, and the value has migrated to the scaffolding, the context, the tools, the verification, the memory, and the human checkpoints. Cole Medin's framing of the harness "mattering more than the model" is not a hot take anymore; it is the operating assumption of the people shipping production software, and Anthropic just proved it on its own codebase.

Notice how cleanly the loop people and the teammate people fit together. Nate Jones's central design rule is that a loop is only trustworthy when it pauses for human decisions, his message-drafting loop stops before hitting send. Claude Tag is that principle productized at the org level: it works autonomously across days but comes back when it needs your input, under a logged identity with channel-scoped permissions. The hard problem of agentic AI was never raw capability; it was control, auditability, and knowing where the machine should stop. The most consequential releases today are the ones that answer those governance questions, not the ones with the highest benchmark.

And underneath it all runs the meter. OpenAI is fabricating a custom inference chip and Mistral is advertising "8x lower cost" because the agentic workflows everyone is racing to build are voracious token consumers. Chase AI's loop engineering explicitly bakes in stop-conditions ("eight iterations, then halt") because, as he flatly notes, "AI isn't free." The same week we learn AI writes most of Anthropic's code, we learn the industry's frontier labs are bending silicon to make that code cheaper to run. Capability and cost are now the same conversation.

The Bigger Picture

Step back and today reads like the moment AI stopped being a product category and started becoming infrastructure, with all the boring, load-bearing concerns that implies: identity, permissions, audit logs, unit economics, and org-chart placement. Claude Tag does not ask to be a smarter chatbot; it asks to be onboarded like an employee, with an account and a scope of authority. That is a different kind of adoption. It means the binding constraint on AI's enterprise spread is no longer model quality, it is trust architecture, and the vendors who win the next phase will be the ones who make delegation safe and accountable, not the ones who top a leaderboard.

The SignalFire data is the human counterweight, and it complicates the doom narrative in a specific way. AI did not thin out engineers; it made them "a lot more productive" with "endless work to do." That is exactly what you would expect if Medin and Jones are right that the job is shifting from typing code to specifying, supervising, and verifying systems of agents. The work did not vanish, it moved up a level of abstraction, toward judgment, taste, and the design of harnesses and loops. Dylan Field's "be at the tails, not the middle" is the same lesson aimed at creative work: when AI floods the median, the durable human value is whatever is out of distribution.

Even Midjourney's wild swing at healthcare, a profitable, investor-free lab proposing an "ultrasonic CT" to make full-body scans a 25,000-square-foot spa amenity, fits the pattern. It is a bet that the scarce thing is not raw model capability but the system around it: hardware, access, regulatory clearance, and a workflow people will actually tolerate. The physics critics are right that ultrasound cannot see through air or bone, and there is no FDA clearance yet. But the framing is the tell. Across the most serious and the most speculative stories today, the message is identical: the model is the easy part. The system is the whole game.

What to Watch

• Whether "65% of code" becomes a hiring story. If Anthropic's internal figure is real and replicable, watch how fast enterprises restructure engineering teams around supervising agents, and whether the SignalFire resilience holds or starts bending for junior roles the data did not break out.
• Custom-silicon economics. Jalapeño is in testing with no ship date. The metric that matters is whether OpenAI's performance-per-watt claim survives production, because cheaper inference is what makes always-on teammates like Claude Tag financially sane to run at scale.
• The governance layer. Claude Tag's per-channel permissions and logged identity are a template. Watch whether "the agent has its own account and audit trail" becomes the baseline enterprise requirement, the way SSO and SOC 2 did.

Go Deeper

• Tag Claude In, Right Where You Already Work — Anthropic's own walkthrough, with the Drew-and-Priya scenario showing Claude tracking a thread, opening a PR, and editing a doc under its own logged account.
• Google's Masterclass on Agentic Engineering — Cole Medin's breakdown of the harness, the vibe-coding-to-agentic-engineering spectrum, and the token economics of CapEx vs. OpEx coding.
• Every Level of Claude Agentic OS — the four-level model (skills/loops, memory, interface, distribution) and why most value lives in the bottom two, including Karpathy's knowledge-base structure.
• The Four-Step Process to Loop Engineer Anything — the four phases of a loop and why the success criteria, not the prompt, decides whether a task should be a loop at all.
• I Stopped Prompting AI One Task at a Time — Nate B. Jones on the loop of loops, with the rule that self-organizing agents are only trustworthy when they stop for human decisions.
• Midjourney Wants to Delete 30% of All Death — Fireship on the "ultrasonic CT," the Midjourney Spa, and the physics-based skepticism the device has already drawn.