Task-success versus tool count and prompt size for small local models
Parent: Mac local LLMs: Agent clients, context and compaction · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Anthropic's tool-writing guide says too many or overlapping tools distract agents and recommends consolidating chained steps into one tool.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Anthropic's tool-writing guide says too many or overlapping tools distract agents and recommends consolidating chained steps into one tool. [source]
- The GitButler local-model gauntlet reports pass rates (Qwen3.5 35B 118/132, Qwen3-Coder 30B 114/132, GLM-4.7-Flash 99/132) on 22 problems, but it does not vary tool count or prompt size, so it is not evidence for this concept. [source]
- A controlled run varying tools (5, 20, 90) and listing size on one local model at fixed task set; not found in any source. [source]
- Anthropic's guide to writing tools states that too many tools or overlapping tools can distract agents from efficient strategies [source]
- GitButler's local benchmark gives Qwen3.5 35B 118/132, Qwen3-Coder 30B 114/132 and GLM-4.7-Flash 99/132 passes on a 22-problem set and does not vary tool count or prompt size [source]
- No source read measures local-model task success against tool count or listing size; the effect for 9B to 35B models is unmeasured here [source]
Children
- No children recorded.