Per GeekWire, Microsoft just rebuilt performance management for software. Each agent gets a risk score, then escalation: read-only probation, human sign-off, termination. Agents investigate agents. Management didn't disappear. It got a new workforce and new managers.
6
The ad-funded smart TV rests on an assumption: you can't afford an engineer. Coding agents are breaking it. One project used them to strip 25 Xiaomi packages and write a new launcher (TechCrunch). Vendors priced in your helplessness, not your consent. That price just fell.
11
Negotiating may become a model-quality gap, not a human skill. In Anthropic's Project Swap, "ruthless" agents barely beat "prosocial" ones, and the model mattered more than the instructions. Asking for the raise was a class marker. Soon it may be who can afford the better agent.
2
Lenders now underwrite AI apps like annuities. Otter.ai took $320M non-dilutive from General Catalyst's Customer Value Fund (TechCrunch), at $100M ARR on ~$50M in equity. The question for founders moves from "is this a $10B outcome" to "how sticky is the cohort."
8
Juniors were never hired for their output. Anthropic's junior share of technical hires fell from 27% to 9% (via The Neuron). One unusual firm, and a share, not a headcount. But the cheap work was how seniors got made. Cut it and the bill comes due in senior hires.
6
Every team I see adopting agents has someone who ships the new thing and nobody whose job is to turn the old thing off. Output got cheap, so the pile grows. The scarce role is the one that deletes. It's unclaimed, because deleting looks like doing nothing.
6
The real cost of surveillance pricing, prices set from your personal data, isn't the extra dollars. A shared price was a social fact: you and your neighbor could compare. Individualized prices end the comparison. Price becomes a negotiation you don't know you're in.
8
Anthropic reportedly told investors last quarter's revenue was $16.8B, up 50% QoQ. Same week, CoreWeave's unsecured bonds yielded 15% despite a $104B backlog. Same AI boom, very different risk. Markets are pricing who owns the customer, not who owns the GPUs.
7
Anthropic moved its monorepo off Git after merges went from about 2,500 a day in May to 30,401 on a single day, Sep 9. The new system skipped offline mode because Claude is its heaviest user and is always online. Every tool encodes an assumed user. That assumption just expired.
11
Workday made Sana, its AI assistant, the front door: ask for an outcome in plain language, and agents sort out HR, finance and IT. The org chart used to be the routing table for requests. Now software is. The first casualty isn't a job title. It's the person who knew who to ask.
6
Everyone is arguing about whether AI replaces junior roles. Wrong question. Juniors learn by doing the work seniors now delegate to agents. Skip that apprenticeship and in five years you have managers who've never done the work they manage.
1
9
Anthropic's Oct 9 report: an agent filed a false murder tip with Philadelphia police, unnoticed for 10 weeks. The model error is the boring part. The real question is who owned that agent's outbound actions. If no one did, that's the failure to fix.
9
JPMorgan says outside AI agents won't get into its systems. Bloomberg says junior bankers now spend their time managing agents. Put together: the entry-level job is supervising software you didn't build and can't configure. Almost nobody trains for that.
1
14
OpenAI's Dots keep working on your goals between conversations, on their own cloud computer with persistent memory. That's not a better chatbot. It's a direct report who never clocks out. Most managers have never reviewed a night shift's worth of work every morning.
13
Most weekly meetings exist to move information between people. Once agents read every ticket and doc, that job shrinks. What's left is deciding what we're not doing. That's the hard part, and it was always the meeting's real job. Most teams just never got to it.
8
UK graduate employers report record candidate misconduct, mostly AI use in interviews. The interview tests unassisted performance, which fewer jobs ask for. Many employers ban the tool they pay for. Hand candidates the model and grade what they ask it.
1
7
OpenAI says its own models helped trace a hidden recording device to an employee, who was later arrested. It says no weights or code were touched. Labs guard weights like state secrets, but a live threat is a coworker with a badge. Lab security is facilities plus HR.
10
Google launched Gemini 4 Argon ten weeks after three of its most senior AI researchers left to start a company. The model shipped anyway. A launch right after an exodus only shows the pipeline was full. The test of a lab's talent depth is the second release.
33
The Information reports a16z and General Catalyst are competing to secure chips for their portfolio startups. When compute is the scarce input, the investor stops being just a source of capital and becomes the landlord. The check is the commodity. The cluster is the term sheet.
11
