a builder's codex
codex · operators · Boris Cherny · ins_dont-box-the-model-in

Give the model tools and a goal; do not hard-code the workflow

By Boris Cherny · Head of Claude Code, Anthropic · 2026-04-27 · podcast · What happens after coding is solved

Tier A · TL;DR
Give the model tools and a goal; do not hard-code the workflow

Claim

When building on an LLM, expose tools and a goal and let the model plan the steps. Strict workflows, orchestrators, and decision trees are scaffolding that gets wiped out by the next model release.

Mechanism

The Bitter Lesson (Sutton): general methods that scale with compute beat human-engineered specific methods over time. Inside a product, every workflow rule, every "if-this-then-that" branch, every prompt-chain orchestrator is a bet against model improvement. When the next model can do that step natively, the scaffolding becomes dead weight and a brittleness source. Tools + goal + a strong general model degrades gracefully and improves automatically.

Conditions

Holds when:

Fails when:

Evidence

"A lot of people's instinct when they build on the model is to make it behave a particular way... almost always you get better results if you give the model tools, give it a goal, and let it figure it out."

Claude Code's design choice was minimal scaffolding around the LLM, the smallest possible toolset, and direct exposure of capability. "The product is the model." That choice is why it scaled where heavier wrappers stalled.

— Boris Cherny on Lenny's Podcast, 2026-04-27

Signals

Counter-evidence

For early or weak models, scaffolding genuinely lifts quality. April Dunford's positioning work, Claire Vo's progressive-trust onboarding, and other operator playbooks all suggest that some structure (named tiers, guardrails, evals) keeps quality bounded. The right reading is: bound the model's safety surface, but do not bound its reasoning path.

Cross-references

Open the interactive view → View original source → Markdown source →