llm-kb
← Back to research
Paper

Community finds retrieval-based MCP tool selection caps at ~23% failure reduction

Practitioners report that agent tool-selection accuracy degrades beyond ~20 MCP tools, and retrieval-based fixes like RAG-MCP only reduce failures by about 23 percentage points (from ~86% to ~43% on one benchmark). A new paper argues the root cause is attention-segment interference, not retrieval quality.

7 engagement·1 source·Mon, Jul 13, 2026, 07:44 AM
Multiple practitioner reports from July 2026 (Chew Loong Nian on Towards AI, DEV Community) converge on the finding that agent tool-selection accuracy visibly degrades when the number of MCP tools exceeds roughly 20. The default fix has been retrieval augmentation (RAG) over the tool catalog, handing the model a filtered short list. RAG-MCP (arXiv:2505.03275) improved accuracy from 13.62% to 43.13% on their benchmark, with a 50%+ reduction in prompt tokens. However, a paper submitted late June (arXiv:2606.16364, "Looking Is Not Picking: An Attention-Segment Account of") argues that retrieval-based fixes cap at about a 23 percentage point improvement because the core problem is attention-segment interference, not retrieval quality. This suggests that further gains require architectural changes beyond better retrieval.

Entities

MCP(concept)RAG-MCP(tool)arXiv:2505.03275(concept)arXiv:2606.16364(concept)Chew Loong Nian(person)

Related

Paper

Researchers identify asymmetric generalization problem in LLM unlearning benchmarks

A new arXiv paper argues that existing machine unlearning benchmarks for LLMs suffer from under-forgetting and over-forgetting due to an asymmetric generalization problem. The authors propose that evaluation must cover diverse query formulations of target facts to reliably measure knowledge removal while preserving unrelated capabilities.

0 engagement·1 source·arxiv
Fri, Jul 10, 2026, 09:31 AM
arXiv
Community

Community discusses which single integration makes agents useful and which extra tool degrades reliability

A Reddit user reflects on the principle that adding more tools to an agent often increases failure modes and review work, while a single narrow integration that solves a real bottleneck yields the biggest improvement. The post asks the community to share which integration made their agent genuinely useful and which tool they removed for harming reliability.

2 engagement·1 source·reddit
Sun, Jul 12, 2026, 07:10 PM
Community

Developer asks community for agent evaluation practices, cites silent breakage

A developer building AI agents reports that prompt or MCP changes often break silently despite passing manual tests. They ask the community about evaluation methods, including fixed test cases, skill-level vs. end-to-end checks, and tools like DeepEval, LangSmith, and Ragas.

10 engagement·1 source·reddit
Sat, Jul 11, 2026, 06:13 PM
Paper

Study analyzes failure trajectories of CLI coding agents as temporal processes

A new arXiv paper presents the first large-scale empirical study of CLI coding-agent failure trajectories, treating failure as a temporal process rather than a final outcome. The study introduces a process-oriented framework to analyze how failures emerge, evolve, and become unrecoverable in LLM-based coding agents.

0 engagement·1 source·arxiv
Fri, Jul 10, 2026, 03:25 PM
arXiv
Community

User reports tool-selection accuracy drops linearly with more MCP servers due to token bloat

A user connected 4 MCP servers to one agent and observed tool-selection accuracy declining linearly with server count. They traced the issue to every tool's name, description, and JSON Schema being serialized into every request, causing token bloat. With 4 servers and ~60 tools, the serialized definitions consumed significant context, degrading performance.

11 engagement·1 source·reddit
Sat, Jul 11, 2026, 08:49 PM