The River Doesn't Wait

Three moves in the last two weeks each handed you the upside and the exposure in the same act: an agent that now lives inside your company's Slack, a Chinese model that does the same work for a fortieth of the price, and a hard reminder that the speed you buy with automation gets paid for in judgment. You can't take the prize without touching the peril.

• Anthropic's Claude Tag puts an agent inside your team's Slack channels as a standing coworker. It's the fastest way yet to hand agents your context, which is also the fastest way to leak it. Ask Meta.
• DeepSeek, a Chinese open-source model, can run an agent's work for about five cents a task where Claude Opus 4.8 runs two dollars. The savings and the security question arrive together.
• Ford rehired 350 veteran engineers it had automated away. The judgment you speed past is the judgment you end up missing.

Here's the thread I'd pull through the last two weeks. The move that actually mattered was agents getting inside the places where your work already happens. And that move hands them the context that makes them powerful in the exact same place it concentrates the risk. The prize and the peril are one thing. You can watch it happen three times over.

It's the same boat from the book: your domain of control, out on the river while the current keeps rising. Two of these moves are surges and the third is the current catching on rock, but navigating any of them this stretch came down to the same thing. You had to read all three settings on the boat at once, because every push carried a catch.

Agents move in, and your context is the prize and the peril

Anthropic shipped Claude Tag, an agent that lives in your team's Slack channels as a standing coworker. You summon it with @Claude, it runs multi-step tasks on its own, and unlike the old bots it's "multiplayer": the whole channel shares one Claude that holds the context and never needs re-briefing. How real is this? Anthropic has been running an internal version of Claude Tag on its own team, and says it now writes 65% of the product team's code. It is building the product with the product.

Here's why it's the story and not just a feature. Put an agent where your people actually talk, and you crack the hardest problem in making agents useful: context. It soaks up how your team really works, the decisions and the reasons behind them, without anyone stopping to feed it. That's the context tension from the book (Chapter 5), and it's the whole game. The context you hand an agent is what turns it from a clever demo into something that acts like a colleague.

It's also the same move that exposes you. Even if Claude Tag captures only the meaning of a conversation rather than every keystroke, that meaning now sits in a system, and systems spread. Meta just ran the cautionary version: its employee-tracking program, the Model Capability Initiative, pooled private conversations, performance data, and transcriptions, then unexpectedly made them visible to the entire company. "We're pausing it while we investigate," Meta said. The context that makes the agent valuable is the context that leaks. Prize and peril, one move.

The cost floor drops, and it's Chinese

The second move is about money, and it cuts both ways just as cleanly. Agents are expensive because they call a model over and over as they work, and that bill has been climbing fast. So Microsoft is testing DeepSeek, a Chinese open-source model, as a cheaper engine for Copilot Cowork, hosted on its own Azure cloud. The gap is not small. Ina Fried of Axios reported DeepSeek V4 running a task for about five cents where Claude Opus 4.8 runs about two dollars. That's the difference between an agent you can afford to run everywhere and one you ration by the task. Lindy, an AI-assistant startup, already switched, routing around the obvious worry by reaching DeepSeek through an American provider. But even a model you host yourself was still built and trained by a company deeply tied to the Chinese government.

The prize is real: at a fortieth of the cost, every use case the price tag had closed swings back open. So is the peril, and it isn't subtle. Plenty of global enterprises hold hard rules against Chinese technology in anything foundational, and an agentic layer is about as foundational as it gets. This is the same shape as the other two moves: the thing that makes DeepSeek attractive, a rock-bottom price on a capable model, can't be pulled apart from the thing that makes it risky, which is where it came from. Prize and peril, one move, this time with a geopolitical edge.

The reimagine hits rocks

The third move is the one nobody put in a press release. Ford quietly rehired 350 veteran engineers, the "gray beards" it had automated out of its quality process, after the AI-driven inspection systems gave what COO Kumar Galhotra called "disappointing results." His VP of vehicle hardware engineering, Charles Poon, put it plainly: "Mistakenly we thought that by just introducing artificial intelligence... that would produce a high-quality product." The veterans are back to train the juniors and fix the tools, and Ford ties the move to real money: CEO Jim Farley credited it with hundreds of millions of dollars in avoided warranty and recall costs.

Reimagining the work was the right instinct. Ford just learned which part it couldn't hand over yet. And there's a name for the bill it paid: a RealClearMarkets essay by Amyn Jan and Mitzi Perdue calls it cognitive debt, the quiet cost that builds as AI does more of the thinking and, in their words, "the organization becomes faster while its reservoir of judgment grows more slowly." It's the cousin of the coherence debt I write about in the book (Chapter 6). The speed is the prize. The thinning judgment underneath is the peril, and it's the same move: every task you let the AI think through is a task your people don't. Ford is what it looks like when that debt comes due.

What I'm watching next

Three things on my list. Whether Microsoft actually ships DeepSeek inside Copilot Cowork, and how many enterprises decide the savings are worth the questions. Whether agents-in-your-tools like Claude Tag become the default, and whether one of them produces the internal leak Meta rehearsed. And whether more Ford moments surface: companies quietly buying back the judgment they automated away.

Sources

1. Anthropic, Introducing Claude Tag
2. Meta pauses its employee-tracking program (Engadget)
3. Microsoft is testing DeepSeek for Copilot Cowork (Axios, Ina Fried)
4. Ford rehires veteran engineers (TechCrunch)
5. Cognitive debt (RealClearMarkets, Amyn Jan and Mitzi Perdue)