Apple ML study compares Retrieve-then-Plan vs Plan-then-Retrieve for LLM agent tool discovery
Apple ML Research · rss · 2026-10-15
Apple ML Research published a scale study of tool discovery architectures for tool-augmented LLM agents. The dominant Retrieve-then-Plan (RtP) paradigm embeds the user query and retrieves similar tools via approximate nearest-neighbor search from large API registries, since finite context windows can't fit all tool definitions.
The study evaluates an alternative Plan-then-Retrieve (PtR) paradigm: the LLM first generates Abstract Tool Stubs — hypothesized function signatures describing the capabilities it needs — which then serve as retrieval queries against the tool registry.
The paper benchmarks both architectures across scales, offering empirical guidance for engineering tool-use in agents.
More from coding & agent
- Turn charts, logos and photos into animated videos in 5 minutes with Claude Code + Remotion — 6 copy-paste prompts — TawohAwa · 2026-10-16
- Enterprise AI agents on 20-year-old healthcare billing data: CPTO went from skeptic to user in 6 weeks — alex_verem · 2026-10-16
- Anthropic eng lead Felix Rieseberg's Claude tutorial distilled into 7 prompts — alex_verem · 2026-10-16
- LangChain Guide: How Madrigal, Abridge and Vizient Scale Agents in Healthcare — LangChain · 2026-10-16
- Full Masterclass: Building an AI GTM Army That Runs Your Sales Pipeline — VibeMarketer_ · 2026-10-16
- Beam Scores 28 on Coding Agent Index, 40 Behind Leaders but Among Most Token-Frugal — ArtificialAnlys · 2026-10-16