The current Agent SDK adopts a whitelist model (allowedTools) for tools, skills, mcp, and subagent authorization. In scenarios with a large number of skills, tools, and subagents (such as over 2000), the following issues arise:
-
Problem description - Users cannot predict in advance which skills are relevant to a problem - All skills must be loaded into allowedTools at once - This results in system messages containing a large amount of tool metadata during initialization - The context occupies too much space, potentially triggering multiple compression failures
-
Suggested Solution: It is recommended that the SDK supports one of the following mechanisms:
a) Vector preliminary screening: Before sending a query, use vector search to quickly filter a subset of skills related to the user's question. b) Dynamic tool selection: Provide an API to allow dynamic adjustment of allowedTools before querying. c) Tool classification routing: Support grouping skills by functional modules or labels, and automatically select relevant groups
-
Expected benefits: - Reduce token consumption during initialization - Avoid multiple compression failures caused by excessively long contexts - Improve performance and stability in large-scale skill scenarios
The current Agent SDK adopts a whitelist model (allowedTools) for tools, skills, mcp, and subagent authorization. In scenarios with a large number of skills, tools, and subagents (such as over 2000), the following issues arise:
Problem description - Users cannot predict in advance which skills are relevant to a problem - All skills must be loaded into allowedTools at once - This results in system messages containing a large amount of tool metadata during initialization - The context occupies too much space, potentially triggering multiple compression failures
Suggested Solution: It is recommended that the SDK supports one of the following mechanisms:
a) Vector preliminary screening: Before sending a query, use vector search to quickly filter a subset of skills related to the user's question. b) Dynamic tool selection: Provide an API to allow dynamic adjustment of allowedTools before querying. c) Tool classification routing: Support grouping skills by functional modules or labels, and automatically select relevant groups
Expected benefits: - Reduce token consumption during initialization - Avoid multiple compression failures caused by excessively long contexts - Improve performance and stability in large-scale skill scenarios