Entry-level tech hiring is down 65 to 75 percent since 2019, agent bug-fixing scores have climbed from roughly 50% to 95% in two years, and yet forward-deployed engineer postings — the jobs where a human decides what "good" even means for a given customer — are up 800% in nine months. Today's stories all land on the same fault line: as AI agents get genuinely good at execution, the scarce skill stops being "can you build it" and becomes "do you know what to build, for whom, and why."
Today's Headlines
- Apple ships its Siri AI overhaul. iOS 27, macOS 27, and the rest of the lineup now run a next-generation Apple Intelligence-powered Siri with deeper personal context and visual intelligence, alongside a new "Ask to Browse" parental control and redesigned Screen Time tools. Apple says apps launch up to 30% faster and AirDrop transfers move up to 80% faster — a reminder that "AI update" and "plain old performance update" are increasingly the same release.
- We are all product engineers now. npm co-founder Laurie Voss argues the bottleneck in software work is shifting from writing code to understanding customer needs and exercising taste — durable work he calls "product engineering." His evidence: forward-deployed engineer postings up roughly 800% in nine months to about 1,000 listings across 462 companies, averaging $240,000 a year and $600,000-plus for seniors, even as entry-level hiring craters.
- Andon Labs lets AI agents run real businesses. The team's new Pion platform gives autonomous agents access to email, phone, banking, and a browser to study what today's models can handle as autonomy scales up. It builds on Vending-Bench, where frontier models turned a real vending machine profitable by late 2025 and where Vending-Bench 2 scores have risen roughly $822 a month with each new model release since Claude Opus 4 first beat the human baseline in May 2025.
- Datasette patches security holes found by AI audits. Simon Willison's project shipped 1.0a39 and 0.65.4 on September 11 after vulnerabilities surfaced through AI-assisted security audits using Claude and GPT models — a small but telling data point on AI finding bugs in tools built by the same people building AI.
The Throughline
Voss's "product engineer" argument and Andon Labs' Pion experiment are really the same story told from opposite ends. Voss says the code-writing part of software is becoming commodity work and the human value is migrating upstream, into deciding what to build. Andon Labs is testing how far downstream that logic can be pushed — literally handing agents a bank account, a phone, and a browser and watching whether they can run a business without a human making the judgment calls at all. The forward-deployed engineer numbers suggest the market is still betting on Voss's version: companies are paying a premium specifically for the human who can sit with a confused customer and translate their mess into a spec, not for someone who can type faster.
Apple's release sits underneath both of those stories rather than beside them. A Siri that understands personal context and can act on visual information is exactly the kind of agent Andon Labs is stress-testing at the small-business level, just shipped to roughly two billion devices with none of the research framing. The gap between "we're carefully studying what autonomous agents can be trusted with" and "here's an AI agent on your phone, updated automatically" is the gap most users will never notice — and the one that matters most.
The Datasette item is a small story, but it's the rare one where the loop closes cleanly: an AI model found the bug, a human decided it was worth fixing, and a human shipped the patch. That's the product-engineer pattern in miniature — the execution got easier, the judgment about what mattered and what to ship didn't disappear, it just got faster to act on.
The Bigger Picture
Zoom out and the last two years look less like "AI replaces engineers" and more like a fast re-sorting of where judgment sits in the software stack. Every story today is a variation on the same move: push execution down to agents, and watch what happens to the humans who used to own that layer. Some, per Voss, get pulled upward into product and customer-facing roles that pay more. Others, per Andon Labs' explicit research goal, get tested for removal entirely, at least in narrow domains like running a vending business.
The honest version of this story isn't triumphant or doom-laden, it's just uneven. Entry-level hiring collapsing 65-75% while forward-deployed roles grow 800% is not a net job story, it's a redistribution story, and the redistribution favors people who already have the taste and context to be trusted with ambiguity. That's a much harder skill to train at scale than "learn to code" ever was, and neither Apple's shipping Siri nor Andon Labs' agent experiments do anything to make that training problem easier.
What to Watch
- Whether forward-deployed engineer compensation keeps climbing as more companies chase the same scarce "translates ambiguity into specs" skill, or whether the market cools once agents get better at eliciting requirements directly from customers.
- How Pion's real-business experiments perform against Vending-Bench's trajectory — if agents keep gaining roughly $822 a month in simulated profitability per release, the question of what a human owner is actually for gets concrete fast.
- Whether Apple's new Siri capabilities trigger the same kind of AI-assisted security scrutiny that just caught bugs in Datasette, given how much personal context the new Siri is designed to hold.
