Busy week. Ten items below. The Uber story is the one I'd put in front of every client, so I'm starting there and working through the rest roughly in order of how much I'd actually change because of it.
Uber burned its whole year's AI budget in four months. Unbounded agent loops, full chat histories re-sent on every turn, no caching, no routing, no spend caps. They've since capped engineers at around $1,500 each. This is the thing I've been warning people about, and now there's a name-brand example. An agent left to loop is a taxi with the meter running and nobody in the back seat. If you're putting agents into production without step limits, caching, and spend alerts, the budget is the first thing that breaks, and it breaks quietly until the bill shows up. Do the cost controls first. It's so much cheaper than doing them after the invoice.
Anthropic shipped Claude Tag, and agents finally get their own identity. Instead of borrowing a developer's API key, an agent gets its own team-wide identity with real permissions, and Workload Identity Federation went GA so the static key — the long secret somebody pastes into a script and then leaks — is now optional. I like this a lot. An agent on a human's key is an audit black hole: you can't say who did what, and when the key leaks, everything it touched is exposed. This is the foundation I'd want under any multi-agent system I build for a client, so it's good to have it as a first-class feature instead of something I bolt on.
LangChain added prompt caching for long-lived Deep Agents. Keep the system prompt, tool definitions, and history warm instead of paying full price every turn. Read this one right after the Uber item, because caching is the laziest, highest-return move in the whole stack. The "up to 80% savings" figure is a vendor number — your real number depends on your prompt — but the floor is "meaningful money for almost no work." If you're building agents and not caching, you're choosing to overpay.
Cloudflare did two genuinely useful things. Temporary Accounts for Workers let an agent spin up an isolated sandbox, run its code, run the tests, and tear the whole thing down in milliseconds. And Cloudflare Workflows got native saga-style rollbacks for durable multi-step transactions. The sandbox answers a question you should already be nervous about: where does the agent run code it just wrote? The answer is never "on a shared dev box." A throwaway sandbox per run, every time. And saga rollback is one of those problems that quietly eats weeks of senior-engineer time at every shop I've seen, so letting the platform own that state machine is a real gift. (They also published a deep-dive on hunting an HTTP/2 bug in the Rust hyper library — a nice reminder that somebody still has to do the unglamorous low-level work holding the stack up.)
OpenAI's quietest release was its best one. Past the goofy "Codex-Maxxing" name is real guidance on token discipline for multi-hour autonomous runs, plus a pattern for connecting private internal MCP servers to frontier models without exposing them to the public internet. That second part matters more than it sounds. The single most common objection I hear from enterprises is "are you going to ship my proprietary code to someone else's cloud?" Being able to stand up private MCP servers against a client's own code and docs, behind their own firewall, is a far more useful capability than another demo of an agent writing a game.
Vercel's AI SDK 7 teaches agents to respect your design system. The release has proper agent primitives, durable workflows, and guaranteed structured outputs, plus a writeup on training coding agents against design-system rules. That last bit is the part I care about — feed an agent your component library and your constraints up front and it builds inside the lines instead of inventing a fifth shade of blue. The "cut frontend overhead 75%" number floating around is decoration; ignore the percentage, keep the technique.
OpenAI and Broadcom unveiled a custom inference chip called Jalapeño. It's an ASIC tuned for low-latency, high-concurrency model execution. Latency is the silent killer of agent loops — when a job is ten sequential steps and each one stalls for ten seconds, the whole thing is too slow to ship, a screensaver with extra steps — so purpose-built silicon aimed at exactly that is worth watching. Watching, not building on. You can't buy a Jalapeño and you won't architect on one this quarter. File it under "the bottleneck is dissolving," and plan for a near future where real-time multi-agent loops are cheap. Don't budget for it today.
Netflix deleted a pile of its own custom code. They moved off a home-grown batch scheduler to Kueue, the CNCF-standard Kubernetes job queue, with zero disruption, and built a real-time service topology map while they were at it. The headline isn't Kueue. It's that the clever custom scheduler somebody was proud of in 2019 is now a maintenance liability, and the boring open-source standard does the job. If one of your clients is still babysitting a bespoke queue manager because a former staff engineer wanted to build one, that's a good thing to rip out — cheaper to maintain, less lock-in, and you can hire people who already know how it works.
Stripe and AWS published the grown-up way to ship AI in a regulated domain. Their compliance-verification agents on Bedrock are built around a strict human-in-the-loop protocol: the model does the high-volume research and categorization, a human makes the final call. This is the exact pitch for a risk-averse client. You're not selling them a robot that approves transactions and hopes for the best. You're selling a tireless research assistant that hands a person a structured brief to sign off on — same leverage, none of the "we let the AI decide and got sued" exposure.
If I had to spend my own week on a subset of this, it'd be the cost and identity items — Uber, Claude Tag, caching, sandboxes. None of it is the fun part. It's also where the money and the risk actually live right now, which is usually a sign of where the work is.
About Heavy Chain
Jason Vertrees is the founder of Heavy Chain Engineering — disciplined software-delivery modernization for the AI era. Heavy Chain helps lower middle-market vertical SaaS companies and PE firms turn scattered AI usage into measurable delivery leverage: 85% faster feature velocity, six-to-eight-week projects shipped in days. If you want help building an AI-native engineering org, book an AI Delivery Assessment.