Boris Cherny's Claude Code Setup #4: The model just wants to use tools

Final entry in the Cherny Playbook — from origin story to design philosophy.

---

The interview

"The model just wants to use tools. That's what I realized."

"Agentic search… it's a fancy word for glob and grep."

"Nowadays I ship 20–30 PRs every day, but every PR is just completely different."

"It just one shotted it… this was my first field AGI moment at Anthropic."

— Boris Cherny on the Pragmatic Engineer podcast with Gergely Orosz, 2026

---

What we heard

This interview surfaces the design principle behind Claude Code that most agent frameworks in 2026 still miss. Three things:

1. "The model wants to use tools." This is a claim about model psychology, not framework capability. Modern Claude models — when given access to tools — default to action. They'd rather run ls than ask what files exist. They'd rather grep than hallucinate a symbol. Systems that force the model to reason in a vacuum before acting underutilize this.

2. "Agentic search is a fancy word for glob and grep." Cherny is being deliberate here. He's not saying "don't build retrieval systems." He's saying: the best retrieval primitive for code-shaped problems is often already installed on your machine. Vector databases have a place. So does rg. Know which one you need.

3. "20–30 PRs a day, each completely different." The throughput figure is secondary — the real claim is variety. Cherny is doing many small, heterogeneous things per day, not one large thing. That shape of work is what an agent loop supports; it's not what a traditional "do the feature end-to-end" IDE supports.

---

What we actually do with this

We design every agent system around a tool-first principle: give the model tools that do real things in the real environment, not abstractions that wrap or simulate the environment.

Our hierarchy of tool preferences, highest to lowest:

| Preference | Example | Why | |---|---|---| | Shell commands | rg, git, curl, ls, jq | Battle-tested, fast, ubiquitous | | Direct API calls | fetch, direct DB query | No wrapper abstraction to debug | | Library calls | @langchain/* utilities | Only when the function genuinely needs an SDK | | Custom abstractions | Hand-rolled wrapper | Last resort; usually a smell |

A common failure mode we see in client engagements: teams build elaborate tool wrappers ("search_codebase", "query_database", "fetch_user") that obscure what's actually happening. The model then has to reason about the wrapper's semantics and the underlying operation. Remove the wrapper, expose the primitive, and the agent gets faster and more reliable.

---

Applied: InterviewLM's tool surface

On InterviewLM, our interviewer persona agents have access to five tools:

1. get_rubric(role_id) — reads structured YAML from disk
2. get_transcript(session_id) — reads current session state
3. score_response(criterion_id, score, reasoning) — writes to session state
4. next_question(category, difficulty) — reads candidate question bank
5. end_session(summary) — closes the session

Four out of five are backed by filesystem operations or simple database reads. One writes structured state. No wrappers, no "smart" retrieval, no elaborate context management. Cache hit rate: 90%+. Tool call success rate: 99.2%. Median latency: sub-50ms per tool call.

The discipline is: every tool we add, we ask "is this the primitive or is this a wrapper?" If it's a wrapper, we justify why the model can't just use the primitive directly. Usually we can't justify it, and we don't add the tool.

---

On "agentic search is glob and grep"

We applied this directly. Our code-understanding agents don't use a vector database. They use rg with structured queries. For a 200K-line codebase, rg returns in under 500ms and the model can read the results directly. The failure modes of rg (it won't find things renamed or semantically equivalent but lexically different) are understood and documented. The failure modes of a vector database for code (false positives on similar-looking-but-wrong code, stale indexes, embedding drift) are not as well understood and harder to debug.

This is the right call for most code retrieval. It is not the right call for natural-language document retrieval (that's where vector search shines). Knowing the shape of your data tells you the shape of the retrieval.

---

The one thing to steal from this

Audit your current agent's tool surface. For each tool, answer: is this a primitive or a wrapper? Every wrapper is a place where the model has to learn your custom semantics instead of the operation's natural ones. Replace three wrappers with primitives this week. Measure the change in tool-call success rate and latency. You will almost certainly see both improve.

---

Series complete — Boris Cherny's Claude Code Setup

Four entries, one workflow, one pattern per entry for production AI work:

1. Origin story: "code is not the bottleneck" — the bottleneck diagnostic
2. 5 parallel Claudes + iTerm2 — the 5-lane pattern with git worktrees
3. CLAUDE.md as postmortem — living memory with a forcing function
4. "The model just wants to use tools" — tool-first design; prefer primitives over wrappers

Next in the broader Playbooks project: Harrison Chase on Production Agents — why better models alone won't ship your agent, the LangGraph runtime, and the Deep Agents harness.