Progressive Disclosure Meeting Notes (Sep 21 2026)

[chughtapan]

Summary

Working group renaming and problem statement review for progressive discovery mechanics

Working Group Logistics: Participants agreed to rename the repository to progressive disclosure.

Problem Statement Review: Draft problem statement detailed developer friction and operational costs from large tool catalogs.

Tool Search Limitations: Tool search implementations fall short due to cache-busting issues and increased latency.

Decisions Aligned

Adoption of progressive disclosure working group name The participants mutually agreed to name the working group progressive disclosure.

Next steps

[Tapan Chugh] Setup Meeting Logistics: Create a calendar invite for the working group and update the meeting entry on meet.modelcontextprotocol.io.
[Tapan Chugh] Coordinate Invites: Start a thread on the Discord group to collect member emails for the calendar invite.
[Tapan Chugh] Update Charter: Update the charter on the model context spec to reflect the group's status as an interest group.
[Kurtis Van Gent] Rename Discord Group: Talk to Danny and D3 about renaming the Discord group.
[The group] Review Documents: Review the problem statement and tool search documents and leave comments by Wednesday.
[Kurtis Van Gent] Refine Document: Clean up the problem statement document incorporating group feedback.

Details

1. Working Group Facilitation and Naming Logistics: Tapan Chugh confirmed they would continue serving as the facilitator for the working group. The participants agreed to rename the working group and repository to "progressive disclosure". Kurtis Van Gent asked Tapan Chugh to create a new calendar invite on modelcontextprotocol.io, update meet.modelcontextprotocol.io with the invite, and update the charter to classify the group as an interest group rather than a working group. Additionally, Kurtis Van Gent noted they would check with Danny and D3 about renaming the Discord group, while Tapan Chugh agreed to start a Discord thread to collect participant emails for the calendar invite.

2. Problem Statement on Large Tool Catalogs and Friction: Kurtis Van Gent presented a draft problem statement detailing the operational costs and developer friction caused by large tool catalogs. They outlined three primary issues: exponential cost increases due to tokens in context windows, degraded model accuracy as choices multiply, and increased latency affecting time to first token. Because the Model Context Protocol lacks a built-in discovery mechanism, developers are forced to manually curate and manage MCP servers on a project-by-project basis, introducing significant friction compared to systems like the G-Cloud command-line interface and skills which feature progressive discovery.

3. Problem Statement on Protocol Scope and Desired Outcomes: Kurtis Van Gent and Luca Chang discussed how command-line interfaces and skills utilize progressive discovery, noting that major AI labs train models on environments like the GitHub command-line interface, though snapshot changes create gaps that skills help address. Kurtis Van Gent argued that progressive discovery must be implemented at the protocol level in Model Context Protocol because servers understand their capabilities and semantics, while clients understand user intent and model limits. The desired outcomes include a bounded entry point, selected expansion for tools, prompts, and resources, server-provided semantics, dynamic authorization and configuration support, and execution method agnosticism compatible with code mode.

4. Feedback on Code Mode, Agent Extensions, and Review Timelines: Luca Chang noted that nesting code mode when both servers and clients use code execution patterns can cause performance conflicts, suggesting potential solutions like server variants. Madhavi Pasumarthi and Luca Chang highlighted parallel work in the agents working group where servers expose agent definitions to group tool sets and reduce catalog bloat. Vijaydeep Sinha described their work on code mode using open API specifications, and Kurtis Van Gent affirmed that progressive disclosure mechanisms must remain compatible with code mode without dictating harness implementations. Kurtis Van Gent established a deadline for feedback on the problem statement document by Wednesday, with cleanup planned for Thursday or Friday and a core maintainer review requested for the following week.

5. Debate on Document Comparisons and Native Function Calling: Tapan Chugh and Kurtis Van Gent debated whether comparisons to command-line interfaces and skills should remain in the problem statement document, with Kurtis Van Gent defending their inclusion due to frequent community discussions. Pedro Rodrigues questioned whether existing client and server solutions like code mode are sufficient, prompting Kurtis Van Gent to explain why tool search alone is inadequate. Reviewing the history of native function calling, Kurtis Van Gent outlined its advantages—such as schema validation, selection controls, parallel calls, and correlation between calls and results—which set a performance baseline that search mechanisms must meet

6. Limitations of Tool Search Implementations: Kurtis Van Gent detailed why various implementations of tool search fall short of solving progressive discovery. Implementing search and execute tools directly on an MCP server causes a loss of native tool calling benefits, while proxy servers introduce hosting overhead and require extra prompt loops. Client-provided tool search implementations (such as those in Claude Code and CodeX) cost an additional model turn and risk cache-busting issues. Furthermore, model-provided tool search suffers from cache-busting problems and inconsistent platform support, as models like Gemini and GLM lack support, and platforms like Anthropic Claude require specific environment variables on Vertex API and Amazon Bedrock.

7. Tool Search Latency, Tool Blindness, and Next Steps: Kurtis Van Gent discussed additional drawbacks of tool search, including increased latency from multi-stage lookups and "tool blindness," where models struggle to discover tools without knowing their names beforehand. Chirag Agrawal raised concerns about whether skills suffer from similar scaling issues if too many exist, to which Kurtis Van Gent responded that skills reduce the problem but would eventually require nested mechanisms. Finally, Kurtis Van Gent encouraged participants to review the documents and leave comments by Wednesday, aiming to finalize the problem statement by next Monday and transition to discussing the grouping repository proposal for skills.