What the Industry Calls Trends, I Call Tuesday

This is not a victory lap. Some of what's lurking out there is a real problem I haven't solved either. But the comparison is worth making plainly, because the gap between what gets funded and what gets built tells you something about both sides.

The Token Bill Reckoning, On Schedule

The research surfaced this stat: Klarna saved sixty million dollars and replaced the workload of 853 employees with an AI agent, then reversed course because complex and emotionally charged queries needed human judgment the agent couldn't supply. JPMorgan runs over 450 agentic use cases in production daily. Uber, Shopify, and GitHub Copilot are all named as major adopters of cloud-routed coding agents, with one report noting enterprises are now "rationing access" because token bills are starting to compete with payroll.

I wrote the cost math in November 2024. Private inference cluster, $766/month, no rate limits, no overages, data goes nowhere. That post was satire because the gap between what I'd priced out and what these companies were about to discover was so wide it read as a joke. It still does. The blueprint was sitting there. They had other plans, and the other plans involved watching usage scale linearly with cost in a way mine never will.

The hardware underneath that math changed once. Nvidia from September 2023 through March 2025, then M4 Max the week Apple announced it, 36GB first, 128GB once the headroom was obvious. One vendor switch, no procurement cycle, no board to convince. The enterprises in this research are still negotiating GPU allocation. I bought a better Mac.

This isn't schadenfreude. It's structural: metered compute makes every agent a recurring decision to keep paying. A box under your desk makes the decision once, at purchase, and then agents are free to fail, retry, and run long without anyone calculating the burn rate. The economics shape the willingness to experiment. Mine let me run six commits in a bad week without thinking about cost. A metered enterprise team has to think about cost on every retry, which means they retry less, which means they ship more brittle automation, which means more of the kind of incident that made Klarna pull back.

Multi-Agent Orchestration, and the Part Nobody Puts in the Deck

Gartner reports a 1,445% surge in multi-agent system inquiries between Q1 2024 and Q2 2025. The framing everywhere is orchestration: a puppeteer agent coordinating specialist sub-agents, modular and clean, like an orchestra. Fountain's recruiting case study gets cited as proof, 50% faster screening using hierarchical orchestration. Gartner forecasts 70% of multiagent systems will contain narrowly specialized agents by 2027, which the same report admits raises "the potential for compounding errors." That's 1,445 teams about to discover what I found in week one.

That admission is the whole story, buried in a sentence about architecture maturity. I didn't read about compounding errors. I lived eight commits of them across two weeks in June. Instructions ignored. Schema mismatches. Silent looping. An undefined constant referenced three times before anyone, meaning me, caught it. Every failure required a human to identify, diagnose, and resolve. The agent was not autonomous. It was not close.

This is the part the orchestration narrative skips: specialization doesn't remove failure modes, it relocates them. A puppeteer coordinating five narrow agents has five new handoff points for instructions to get dropped instead of one. The industry is optimistic because the case studies that get published are the ones that worked. I publish the ones that didn't, because that's the only honest account of running an agent against real, messy, seven-app production code week to week.

Code Agents Are the One Place the Industry and I Actually Agree

This is the category where the research and my experience line up almost exactly. Claude Code writes about 4% of all public GitHub commits now. Senior developers, ten or more years of experience, pick it over Copilot 46% to 9% in JetBrains' survey. Developers spend more hours reviewing AI-generated code than writing new code: 11.4 versus 9.8 hours a week, a reversal of the old pattern.

That reversal is exactly my week. Ghost routes. A Docker seed file stale for five weeks. Twenty inbox0 endpoints quietly 404ing since the day they were written, caught only because I went looking. None of that gets caught by an agent reviewing its own output uncritically. It gets caught by a human reading what the agent produced and asking whether it's actually true. "The job becoming orchestration" is a nicer way of saying the job became code review, full time, for everyone. I'd add one thing the surveys don't capture: the review burden doesn't shrink as the agent gets better. It just moves to higher-stakes mistakes, the kind buried in a security audit rather than a syntax error. My thirteen closed findings in DrawmaticAI, the IDOR on every diagram endpoint, the JWT secret regenerating on restart, none were caught by tests. They were caught by treating the agent's confidence as a thing to verify, not a thing to trust.

Data Sovereignty as a Market Category, Finally

The research flagged outcome-based business models and FinOps for agents as emerging trends, the economics of agent deployment are about to get formalized into something companies budget for and govern. Good. Someone should. My no-business post made the same argument on June 4th about a narrower slice: Notion AI, Reflect, and BeforeSunset all want your life synced into their infrastructure so their models can hand it back to you summarized. The ninth app in this suite does that same weekly review, across meetings, email, code, and tasks, and your data never moves.

The macro trend is enterprises asking what agent infrastructure costs them in dollars. My version of that question was asked at the level of a single person's calendar and inbox, which is a smaller stakes version of the same problem: who owns the substrate the agent runs on, and what happens to your information while it's reasoning about you. The dollar question is catching up to a question about ownership that was always more fundamental. Enterprises are starting to notice both at once.

Where the Industry Is Ahead, and I Should Say So

Voice as a real bidirectional channel, with proper accent and multilingual handling, isn't something I've built and it isn't something I was early on. Neither is the AI co-scientist work, DeepMind's Co-Scientist and FutureHouse's Robin generating actual hypotheses for drug repurposing. That's a different scale of problem than seven local apps and a steward who watches for me. Worth naming honestly: some of what's lurking out there is lurking because it's genuinely hard, not because the industry hasn't caught up to something I already solved.

The verifiability gap will close for browser agents the way it has for code, because that's the bottleneck holding back the next obvious wave of automation.

What I'd put money on next: the verifiability gap closing somewhat for browser agents, the way it already has for code, because that's the bottleneck holding back the next obvious wave of automation. And a wave of quietly cancelled pilots once the ROI scrutiny the research flagged actually lands, the same way Klarna's reversal landed. I won't be surprised when that shows up as a journey post on someone else's site, eighteen months after it shows up as a commit on mine.