---
source: https://iinventedlight.com/library/notes/ao-tools-llms-as-tool-makers.md
title: "LLMs as tool makers: build once, reuse cheaply"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2305.17126] Large Language Models as Tool Makers, CC BY 4.0; arXiv, [2308.11432] A Survey on Large Language Model based Autonomous Agents, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was run; only the fetched arXiv abstract pages were read, so results are as reported by the authors."
category: agent-operations
topic_area: "agent-ops: tools"
tags: ['agent-operations', 'tools', 'cost']
sources:
  - id: S1
    title: "[2305.17126] Large Language Models as Tool Makers"
    publisher: arXiv
    url: https://arxiv.org/abs/2305.17126
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2308.11432] A Survey on Large Language Model based Autonomous Agents"
    publisher: arXiv
    url: https://arxiv.org/abs/2308.11432
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# LLMs as tool makers: build once, reuse cheaply

**Answer:** In the LATM framework, a strong (expensive) model writes reusable tools and a lightweight model uses them, so the one-off cost of making a tool is spread across many cheaper calls.

## Two phases

1. **Tool making:** a capable LLM crafts a tool (for example a small Python function) for a family of tasks.
2. **Tool using:** a separate, cheaper LLM applies the cached tool to new requests.

On the serving side, tools are generated and cached as new kinds of request arrive, and later requests reach them through APIs.

## Why it saves money

The authors assign tool making to a powerful model because it needs more sophisticated reasoning, and give tool use to a lightweight model. They report that this division of labor significantly reduces average cost while keeping strong performance.

## Context

A survey of LLM-based autonomous agents (S2) describes the construction of such agents in a unified framework; tool use is one of the abilities that framework covers. LATM adds the idea that the agent can extend its own toolset.

## Operator notes

- Cached tools need review and tests: a bad tool gets reused many times.
- Works best when many requests share a structure.

## Claims
- id: C1
  claim: LATM is a closed-loop framework where LLMs create their own reusable tools.
  source: S1
  evidence: "LLMs create their own reusable tools for problem-solving"
- id: C2
  claim: Another LLM acts as the tool user that applies the built tool.
  source: S1
  evidence: "another LLM acts as the tool user, which applies the tool built by the tool"
- id: C3
  claim: Tool making is assigned to a powerful but resource-intensive model.
  source: S1
  evidence: "we assign this task to a powerful, albeit resource-intensive, model"
- id: C4
  claim: The simpler tool-using phase is delegated to a lightweight model.
  source: S1
  evidence: "the simpler tool-using phase is delegated to a lightweight model"
- id: C5
  claim: The survey reviews the construction of LLM-based autonomous agents.
  source: S2
  evidence: "we first discuss the construction of LLM-based autonomous agents"

## Sources
- [S1] [2305.17126] Large Language Models as Tool Makers. arXiv. https://arxiv.org/abs/2305.17126 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2308.11432] A Survey on Large Language Model based Autonomous Agents. arXiv. https://arxiv.org/abs/2308.11432 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).