the token diet works; the trillion-dollar promises don't reconcile
Week ending January 16, 2026; archive coverage: January 13.
This was a week where the tooling got genuinely smarter and the money got genuinely stranger. The engineers shipped things that respect your context window like it's an actual budget. The finance people announced numbers with more zeros than definitions. And somewhere between the two, the uncomfortable question kept surfacing: does any of this actually move delivery?
My read: the practitioner layer is maturing fast, and the hype layer is aging badly. Those aren't contradictory. They're the same story told at different altitudes.
context is finally treated as a resource
The thing I hit constantly is MCP schema bloat. Load six servers, sixty tools, and you've spent ~47,000 tokens on tool definitions before the model reasons about anything. That's a third of your effective context gone to configuration. Schmid's MCP CLI inverts it: a single Bun binary that does dynamic discovery, so the agent asks "what servers exist?" and "what are the params for tool X?" only when it needs them. He clocks the same scenario at ~400 tokens. A 99% cut.
The catch is that dynamic discovery adds round trips and depends on the agent being disciplined about pulling only what it needs. You're trading upfront cost for iterative cost, plus a daemon with a 60-second idle timeout to hide the spawn latency. But this is the right trade, and it's the kind of unsexy plumbing that beats framework gravity every time. Tool filtering via allow/deny globs is the quiet bonus.
typed browsing that shows its work
Vercel's agent-browser is the same philosophy pointed at the browser: a native Rust CLI that hands agents an accessibility-tree snapshot with stable refs (@e1, @e2), then lets them act by ref. What I like is the honesty of the failure modes. When a click is blocked, it tells you the target was covered by <div#consent-banner> and instructs you to dismiss it and re-snapshot. No self-graded demo, no confident lie about success. It also ships content-boundary markers, domain allowlists, and an auth vault so the LLM never sees the password.
The surface area is enormous—network interception, React introspection, MCP profiles, session persistence—which is its own risk. But it's inspectable. You can read exactly what the agent did. That's rarer than it should be.
the part where we ask if any of this pays
So does better tooling produce value? Codemanship's "Great Filter" throws cold water, and I mostly agree. Three years in, no app-store surge, no bottom-line lift, no GDP bump. His reading of the DORA data is the sting: the teams getting gains from AI were already high performers, because they'd already killed their bottlenecks—big merges, after-the-fact testing, PR queues. Bolt a code firehose onto a slow process and you get a marginally slower slow process with more code to review.
The uncomfortable claim is that the capability gap takes years and roughly a fifth of your dev budget to close, and you can't buy it flown in from Silicon Valley. I'd preserve the uncertainty—this is one analyst reading aggregate data—but it matches what I see. AI is a multiplier on your pipeline, and multiplying by a broken pipeline doesn't help.
the numbers nobody can check
Meanwhile, Dave Friedman's teardown of the data-center deals is the most clarifying thing I read. Over half a trillion in announced commitments, and the units aren't standardized. A "gigawatt" might be a press release, a grid interconnect, a built-but-unloaded site, or a sustained load curve—separated by 12–24 months of execution risk. "Up to $100B" is an option, not capex, but the market keeps pricing optionality as commitment. His framing that these announcements are coordination technology—reflexive attempts to make the world conform to the plan—is more interesting than "it's all promotion," and more damning.
the wildcard
And for fun: Sakana's Digital Red Queen evolves Core War warriors with LLMs against a growing history of opponents instead of a static benchmark. Warriors get more generally robust, and independent runs converge in behavior but not in source—function, not genotype. It's a self-contained sandbox where nothing escapes, which is exactly why it's a sane place to study adversarial AI dynamics and automated red-teaming.
what I'd do Monday
Audit your agent's context: how many tokens go to tool schemas before it reasons? Try dynamic discovery. Prefer agent tooling that reports concrete failures over tools that self-grade. And before you credit AI for a productivity win, measure your batch size and lead time—if work sits in queues, AI just fills the queue faster.
vocabulary inflation
Digital Red Queen (DRQ): Sakana's term for LLM-driven adversarial program evolution where agents must keep adapting just to hold their relative fitness. Optionality priced as commitment: Friedman's phrase for treating "up to $100B" headlines as executed capex.