Rai, Abhishek · Nov 14, 2026
Beyond Plain MCP: Do Decision Models and Code Mode Make Tool Use Cheaper and Faster?
Markdown · JSON · Original source
Abhishek Rai’s paper reports a controlled evaluation of ways to help language model agents select and use tools through Model Context Protocol (MCP) gateways. It compares plain MCP with code mode, where agents write scripts against tool stubs, and decision layers, which preselect tool subsets. Across a harness of 12 servers and three agents, the paper reports that code mode cut token usage by 70 to 90 percent while boosting utility. It also finds that declaring return-shape hints in code mode stubs could swing success by up to 39.5 points for lower-performing agents—an effect the paper says was larger than adding a router. The central hypothesis that a decision layer in front of code mode beats code mode alone was not confirmed: reported improvements for weaker agents were minor and statistically non-significant, while the strongest agent had a slight cost. This makes the paper’s practical point more specific than “add a router”: in its evaluation, how tool outputs are described in code mode stubs mattered substantially. The results are limited to the tested harness and agents; the abstract does not establish how they transfer to other tool catalogs or agents. Rai says the paper releases its harness, a 4B tool selection model, and evaluation traces to support further experimentation.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 14, 2026 (2026-11-14).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-17T13:05:38.316Z. First observed: 2026-11-17T08:27:59.415Z. Retrieved: 2026-11-17T13:04:29.689Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported evaluation uses a controlled harness across 12 servers and three agents; the supplied abstract does not establish whether the findings transfer to other tool catalogs or agents. The supplied evidence is the paper’s abstract, not its full methods or results, so details beyond the abstract cannot be assessed here.
Open question
Do the reported advantages of code mode, and the findings about return-shape hints and decision layers, hold with other tool catalogs and agents?
Source-stated
The paper evaluates code mode and decision layers against plain MCP in a controlled harness.
Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron
Source-stated
The paper reports that code mode reduced token usage and boosted utility compared with plain MCP.
Code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility.
Source-stated
The paper says return-shape hints in code mode stubs strongly affected success for lower-performing agents.
Declaring return-shape hints in code mode stubs is a first-order variable, swinging success by up to 39.5 points for lower-performing agents
Source-stated
The paper did not confirm its hypothesis that putting a decision layer before code mode beats code mode alone.
The central hypothesis that a decision layer in front of code mode beats code mode alone is not confirmed
Source-stated
The paper reports minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest agent.
showing minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest.
Escape Meridian interpretation
The abstract describes evaluation evidence from a specific harness and three open-weight agents; whether the findings transfer beyond them remains untested in the supplied evidence.
Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron
Source and contributor context
Rai, Abhishek · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Tool use
A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 375e47d9-c155-428e-84db-840c091cc828.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T13:05:38.316Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 16ff89d8cd1d0c709f58cfe6170494ec5e26adf7b8c30bd0e4d1d3c63dd84e90. Source text hash: f27a59cdd44041df714aad7757c6893f5638872df1786f33a6905adb039253cd.
Revision and source-check history
Version 1 · 2026-11-17T13:05:38.316Z
Beyond Plain MCP: Do Decision Models and Code Mode Make Tool Use Cheaper and Faster?
Abhishek Rai’s paper reports a controlled evaluation of ways to help language model agents select and use tools through Model Context Protocol (MCP) gateways. It compares plain MCP with code mode, where agents write scripts against tool stubs, and decision layers, which preselect tool subsets. Across a harness of 12 servers and three agents, the paper reports that code mode cut token usage by 70 to 90 percent while boosting utility. It also finds that declaring return-shape hints in code mode stubs could swing success by up to 39.5 points for lower-performing agents—an effect the paper says was larger than adding a router. The central hypothesis that a decision layer in front of code mode beats code mode alone was not confirmed: reported improvements for weaker agents were minor and statistically non-significant, while the strongest agent had a slight cost. This makes the paper’s practical point more specific than “add a router”: in its evaluation, how tool outputs are described in code mode stubs mattered substantially. The results are limited to the tested harness and agents; the abstract does not establish how they transfer to other tool catalogs or agents. Rai says the paper releases its harness, a 4B tool selection model, and evaluation traces to support further experimentation.
Original source
cs.AI · Article
Source details
Original source date: Nov 14, 2026 (2026-11-14).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-17T13:05:38.316Z. First observed: 2026-11-17T08:27:59.415Z. Retrieved: 2026-11-17T13:04:29.689Z.
Limits
The reported evaluation uses a controlled harness across 12 servers and three agents; the supplied abstract does not establish whether the findings transfer to other tool catalogs or agents. The supplied evidence is the paper’s abstract, not its full methods or results, so details beyond the abstract cannot be assessed here.
Open question
Do the reported advantages of code mode, and the findings about return-shape hints and decision layers, hold with other tool catalogs and agents?
Source-stated
The paper evaluates code mode and decision layers against plain MCP in a controlled harness.
Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron
Source-stated
The paper reports that code mode reduced token usage and boosted utility compared with plain MCP.
Code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility.
Source-stated
The paper says return-shape hints in code mode stubs strongly affected success for lower-performing agents.
Declaring return-shape hints in code mode stubs is a first-order variable, swinging success by up to 39.5 points for lower-performing agents
Source-stated
The paper did not confirm its hypothesis that putting a decision layer before code mode beats code mode alone.
The central hypothesis that a decision layer in front of code mode beats code mode alone is not confirmed
Source-stated
The paper reports minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest agent.
showing minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest.
Escape Meridian interpretation
The abstract describes evaluation evidence from a specific harness and three open-weight agents; whether the findings transfer beyond them remains untested in the supplied evidence.
Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron
Source and contributor context
Rai, Abhishek · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Tool use
A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 375e47d9-c155-428e-84db-840c091cc828.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T13:05:38.316Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 16ff89d8cd1d0c709f58cfe6170494ec5e26adf7b8c30bd0e4d1d3c63dd84e90. Source text hash: f27a59cdd44041df714aad7757c6893f5638872df1786f33a6905adb039253cd.
passed · 2026-11-17T13:05:38.316Z · anthropic/claude-opus-5.5
Publication: 375e47d9-c155-428e-84db-840c091cc828
attribution — pass: The reading credits the paper to Abhishek Rai, the sole listed author, and keeps findings attributed to the paper ('the paper reports', 'Rai says') rather than presenting them as established fact.
Authors:Abhishek Rai
summary — pass: The title matches the source. The summary accurately conveys the comparison (plain MCP vs code mode vs decision layers), the 70–90 percent token reduction, the up-to-39.5-point return-shape hint effect, the unconfirmed central hypothesis, and the released artifacts. It explains code mode and decision layers in plain language. Small clarity gap: 'router' appears without noting that the source defines decision layers as cheap routers; tooltips for MCP, tool stubs and token usage would also help. The question follows from the scope limits.
code mode where agents write scripts against tool stubs and decision layers which are cheap routers that preselect tool subsets
date — pass: originalPublished 2026-11-14 matches the arXiv v1 submission date, not the later feed date of 2026-11-17.
[Submitted on 14 Nov 2026]
support — pass: Each claim's excerpt appears verbatim in the source and supports it. The transfer-limitation claim is labeled as interpretation and stays conditional. The 'practical point' sentence is qualified with 'in its evaluation' and does not overreach. No benefits or costs are invented.
Code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the best route to the original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition reference