2026-03 — Tom Blomfield on the economics of agentic coding

Why this matters

Tom Blomfield (Monzo founder; now General Partner at Y Combinator) delivered the highest-reach 2026 articulation of the economics-of-AI-coding thesis. The quote that travelled furthest:

“The entire Accenture workforce is about to be outperformed by a 24-year-old who learned Claude Code last Tuesday.”

The line is hyperbolic on its face. What makes it weight-bearing is the observation base behind it: as a YC partner, Blomfield sees hundreds of startups’ inside view. Specifically:

“About a quarter of YC founders told [me] that 95% of their code is written by AI.”

That number — independently corroborated by Anthropic’s own internal numbers, where cofounder Jack Clark said in February 2026 that Claude writes “comfortably the majority” of Anthropic’s code and “could handle nearly all of it by year-end if current progress continues” — is the load-bearing claim. Once you accept that data, Blomfield’s stronger framings (5-engineer teams matching 50-engineer output; software engineering’s existence-shape changing in 6-12 months) become arguable rather than absurd.

Where this fits in this notes set: the rest of the repo is how you make Claude Code do useful work (harness, hooks, plan mode, skills, subagents). Blomfield’s talk is the why is this load-bearing. The economics frame motivates the engineering discipline.

What Blomfield actually said

Across his LinkedIn posts, conference appearances, and the YC content circuit through 2026-Q1, the recurring claims:

  1. “95% AI-written” at the small-startup end. About a quarter of YC founders self-report this to him. The other three- quarters are presumably somewhere lower but rising. The distribution is the interesting thing — the frontier of AI-written-code-proportion is no longer a research demo; it’s already shipping production startups.

  2. Teams of 5 doing what used to take 50. Cursor, Windsurf, Lovable cited as $100M+ revenue companies with “astonishingly small teams.” This isn’t a future claim; it’s a current observation. The leverage already exists.

  3. “Software engineering won’t exist in the same way” — in 6 to 12 months. The qualifier is doing work: the profession may persist; the daily activity of typing-code-into-an-IDE shifts toward something else (specification authoring, review, verification, harness design).

  4. The Accenture line specifically. It’s a barbed framing aimed at the large-services-firm-billing-model — work that’s priced as “humans at scale” is exactly the kind that an individual with Claude Code can underbid. Whether or not the 24-year-old literally outperforms the entire workforce, the unit economics of consulting shift before the technical ceiling of agentic coding does.

Cross-reference: Anthropic’s own self-report

The Blomfield observations land harder because Anthropic itself publicly confirmed adjacent numbers in 2026. From Jack Clark in February 2026, paraphrased across multiple writeups:

This matters because Anthropic isn’t a YC startup with selection bias. It’s a sophisticated AI lab building the agent itself. If they are using Claude Code to write Claude Code’s own code, the regression test on Blomfield’s broader claim is internal.

What this means for someone working with Claude Code

If you accept the economics frame — and the data forces some acceptance even if not the strongest version — the practical consequences:

  1. Your skill investment in CC compounds quickly. The leverage is bigger than the leverage you had with previous dev-productivity tools (IDEs, version control, search). The return on a week spent learning Claude Code well is larger than the equivalent week spent learning Git was in 2010.
  2. Harness design is the moat. If everyone has access to the same model, the differentiator is what surrounds it — the CLAUDE.md, the skills, the hooks, the subagent patterns. See the harness engineering entry for why this is The Argument of 2026.
  3. The market for “agent-assisted” labour is bifurcating. At one end: solo operators and 5-engineer teams reaching $100M ARR by deploying agentic-coding leverage to the hilt. At the other: large services firms (Accenture, Infosys, Capgemini) selling priced-by-headcount engagements that are becoming structurally uncompetitive against the same work done by smaller teams using agents.
  4. Skill churn is real and continuous. What you learned about Claude Code six months ago is partly outdated now (this notes set exists in part to track that). Plan to relearn, not to plateau.

Where Blomfield himself ends up

A useful framing he repeats: the founder skill set is what benefits most from AI coding leverage. Founders historically had to defer specific implementation choices to engineers because typing-out-the-code was a real constraint on time. With that constraint relaxed, founders can prototype, iterate, and ship product changes themselves. The bottleneck moves upstream to what to build, not who can build it.

This is consistent with the YC-portfolio observation: at the seed stage, founders with high taste in product + high comfort with agentic tooling are already outshipping teams that have better engineers but slower iteration cycles.

Try this

Three small experiments to calibrate against the Blomfield claims for your own situation:

  1. Track the ratio. For one week, after each code change you ship, log whether you typed the bulk of it or Claude Code did. After a week, you’ll have a number. If you’re below the YC-startup 95% median, ask why — is it because your codebase is genuinely harder, because your CLAUDE.md isn’t tuned, or because you’re not yet trusting the agent for work it would actually do well?
  2. Time a feature both ways. Pick a small feature. Implement it with CC in plan mode + verification (spec → plan → code with reviews); separately, estimate how long it would have taken you typing the same feature manually. The delta is your leverage number.
  3. Read Blomfield’s posts directly. His LinkedIn feed has the unfiltered version; the press writeups quote the most inflammatory lines but lose the qualifiers.

Sources