The Last Mile Is Code

gc.cpp — 1 million lines of code transpiled from LISP.

Agents got really good at code, really fast.

Represent any sort of problem as code.

SWE Bench — Pro agent is contamination-resistant; quality has moved from 33 in Aug '24 to 94; Pro is 78.

40 → 70 this year.

The Stack Has Changed

The old stack:

programming language --compiler--> machine code

The new stack:

natural language --LLM--> programming language --compiler--> machine code

The hard part isn't the code anymore — it's everything underneath it: ClickOps, infra, API keys/secrets.

Andrej Karpathy — "Building a modern app is a bit like assembling IKEA furniture."

The Last Mile

For most of this, this part has always been someone else's job (a platform team, an SRE rotation, the person on call).

It's wildly lopsided: agents have either what it was trained on, or it's in context.

Put the last mile in code, too.

Take action in code space; a runtime (IaC) reflects it back into the cloud.

IaC is like git diff, but for your infra — see what it's going to do before it does it.

Vibe infra like apps, by asking — code we can review, approve, and merge in code.

+20% — agents do better when they act by writing code, not by emitting config (from CodeAct — Executable Code Actions Elicit Better LLM Agents).

Agents now drive 28% of deployments, up from 4% late last year.

Shipping isn't a separate team's job anymore.