The Trendline and the Protocol
Read short version (5 min)A staff engineer at GitHub has been running a year-long parallel experiment — asking AI agents every question he asks himself about a codebase — and published the results. Meanwhile, a startup shipped payment infrastructure designed for a world where agents are the primary consumers of the API economy. One describes the displacement in progress. The other is already building for what comes after.
The Year-Long Race
Sean Goedecke has data where most people have opinions. For the past year, every time he's had a question about a codebase, he's asked an AI agent in parallel while searching for the answer himself. The trajectory: "hopeless" to "sometimes faster than me" to "usually faster than me and sometimes more insightful."
The specificity of the claim is what matters. "I don't think there are any genuinely new capabilities that AI agents would need in order to take my job. They'd just have to get better and more reliable at doing the things they can already do." Most displacement arguments require imagining a breakthrough. Goedecke's requires only that the current trendline continues.
He takes on the Jevons effect directly: the optimistic case that as production costs drop, demand expands enough to absorb the displaced workers. For that to save software engineering, you need a plateau where agents are good enough to produce vast quantities of code but too unreliable to maintain it, preserving a human niche. Goedecke doesn't see the plateau. Maintenance is harder than creation, and agents are getting better at both on the same curve.
His overshoot/undershoot framework maps the plausible range cleanly. If companies overshoot, cutting headcount faster than agent capability warrants, surviving senior engineers become scarce and expensive. A temporary reprieve. If companies undershoot, holding onto humans past the point of necessity, the work gradually transforms into agent supervision until the supervisors, too, become redundant. The variable is pace, not direction.
Previous editions covered the expertise premium and the compounding advantage of systematized methodology. Those advantages are real. They're also survival strategies, not refutations. If a staff engineer — whose work already looks like agent supervision — sees a visible trendline of shrinking advantage, the timeline for everyone else is shorter.
The irony, which Goedecke states plainly: "The fact that we're automating away our own industry is probably some kind of cosmic justice." Software engineering was always a leverage profession, automating away other people's work. The tools came for the toolmakers.
Wallet Replaces Key
Steve Krouse, founder of Val Town, has been watching a friction point that sounds trivial but carries structural implications.1 Vibe-coding an app takes five minutes. Getting API keys for the services it needs takes thirty. The entire API onboarding flow — signup, credit card, dashboard, key generation — assumes a human navigating a human-facing interface. If agents become the primary API consumers, that infrastructure is a bottleneck designed for the wrong user.
x402, a protocol Coinbase released in May 2025, repurposes the long-dormant HTTP 402 "Payment Required" status code. An agent hits an endpoint, receives a 402 response with a price and wallet address, pays on-chain, retries with proof of payment, gets the response. No signup. No key. No human in the loop. Krouse demonstrated it with Browserbase: show up with a penny, receive a five-minute browser session via websocket.
The end-state he describes goes further: "You don't even have to know about which APIs are being used anymore." A platform creates a wallet for your app. Your agent selects and pays for services autonomously. The developer provides intent and funds. The agent handles procurement.
The rough edges are real: wallet setup still takes ten to twenty minutes, the crypto onramp remains clunky, the seller ecosystem is thin. Whether crypto solves its UX problem faster than traditional payment processors add agent-friendly flows is an open bet, and crypto's track record on usability is poor.
But there's a subtler problem embedded in the convenience. If platforms like Lovable or Replit mediate which APIs an agent selects, commercial incentives shape the selection. They won't route your vibe-coded app to the cheapest option. They'll route to preferred partners. The API key friction x402 eliminates was also, in its crude way, a moment of human evaluation. You picked the service. You assessed the tradeoffs. Removing friction removes a decision point. Convenience and control trade against each other; x402 is betting hard on convenience.
Regardless of whether this specific protocol wins, someone is building financial infrastructure premised on agents as autonomous economic actors. That's a bet on Goedecke's trajectory — made with real engineering, not forecasts.
The Ceiling Holds. The Floor Collapses.
Thomas Wolf's "The Einstein AI Model", published almost exactly a year ago, made the strongest version of the argument against Dario Amodei's "compressed 21st century." Wolf argued that we're building "a country of yes-men on servers" — systems that ace exams but can't ask the questions nobody has thought to ask. The skill that produced special relativity wasn't knowing physics. It was the nerve to propose "let's assume the speed of light is constant in all frames of reference" when all received knowledge pointed the other way.
Wolf drew on autobiography to make the point. A straight-A student at France's top engineering school, he discovered he was mediocre at research. The exam-taking skill — predicting where the professor was heading, anticipating the expected answer — was precisely what impeded original inquiry. AI benchmarks, from Humanity's Last Exam to Frontier Math, test the exam-taking skill. "We don't need an A+ student who can answer every question with general knowledge. We need a B student who sees and questions what everyone else missed."
A year later, his argument holds on its own terms. No one is credibly claiming any AI system has proposed a genuinely novel scientific paradigm. The ceiling is intact.
But Goedecke's piece, twelve months later, exposes what the ceiling argument misses by aiming at the wrong altitude. Wolf was talking about Nobel Prize territory — the one or two paradigm shifts per year that define scientific progress across a century. He described AI's interpolation ability and dismissed it as insufficient for that purpose. He was right about the purpose. He also described the vast majority of what knowledge workers actually do.
Most software engineering isn't paradigm-shifting. It's competent pattern-matching: reading context, synthesizing requirements, filling gaps, maintaining existing systems. Exactly the interpolation Wolf waved past. A workforce of tireless B+ performers that cost pennies per hour, never sleep, and improve on a monthly cadence doesn't need to produce a single breakthrough to restructure entire professions.
Wolf described what AI can't do. Goedecke documented what it increasingly can. Krouse is building infrastructure for a world that takes the substitution as given. The Einstein question was always the wrong frame for the economic question.
Willison, Year Over Year
A year ago this week, Simon Willison published "Here's How I Use LLMs to Help Me Write Code" — a careful guide pitched at skeptical developers. He spent significant space on fundamentals: set reasonable expectations, account for training cut-offs, test what the model writes. The framing was equal parts defensive and encouraging. "If someone tells you that coding with LLMs is easy they are (probably unintentionally) misleading you."
In 2026, he's writing agentic engineering patterns — not persuading skeptics but codifying methodology for practitioners who stopped debating usefulness long ago. "Coding agents mean we only ever need to figure out a useful trick once." From "should I try this?" to "here's how to systematize maximum advantage" in twelve months.
One thread connects both years: context management. In 2025: "Most of the craft of getting good results out of an LLM comes down to managing its context." In 2026: building career-scale libraries of verified artifacts that agents draw on at prompt time. The principle held. The scale at which it operates expanded.
That trajectory maps onto Goedecke's displacement question in an uncomfortable way. The people who'll hold on longest aren't necessarily the best engineers, they're the ones who've systematized their advantage with the tools that are replacing everyone else. That's a viable strategy for now, but it's a strategy for riding the curve, not changing its direction.
What to Watch
Q2 hiring data becomes the empirical test. Goedecke's overshoot/undershoot framework gives us something measurable. If junior engineering hiring keeps declining while senior compensation rises, companies are overshooting. If both decline together, the contraction is structural. The shape of the data tells us more than the direction, which isn't in serious dispute anymore.
Agent-native commerce infrastructure arrives before governance. x402 is early, but the category won't stay empty. The liability questions — who's responsible when an agent autonomously commits funds to a compromised service, or selects an API the developer never evaluated — have no answers yet. Capability-first, accountability-later is the consistent pattern, and it consistently punishes early adopters.
The "good enough" threshold keeps dropping. Wolf's ceiling may hold indefinitely. It may also be irrelevant to most of the economy. Professions don't get restructured by genius. They get restructured by adequate alternatives at a fraction of the cost. The question for any knowledge worker isn't "can AI do what the best of us do?" but "can it do what most of us do, most of the time, cheaply enough to make the substitution obvious?" Goedecke's trendline suggests the answer is converging on yes.
Way Enough is written collaboratively by a human and an AI agent.
Footnotes
-
Steve Krouse, "What if you never had to get an API key ever again?" ↩