Scenario
This is not exactly a bug, and developers can work around it themselves, but it would be better if GA had a native mechanism to control the system prompt/tool prompt size.
As GenericAgent gains more tools, injecting the full tool schema into the prompt on every startup or every request will quickly make the prompt too long, slow down startup, waste tokens, and make the model navigate through many irrelevant tools.
In real use, many tasks only need core tools: file read/write, command execution, file search, messaging, and status queries. Long-tail tools should be discovered and loaded on demand.
Current Pain Points
- The full tool schema goes directly into the prompt, making the initial context heavy.
- Even a simple status query carries the entire tool list.
- The more tools are added, the slower the base experience becomes.
- The model is more likely to choose irrelevant tools because tool selection cost increases.
Suggested Direction
Introduce “core tool warm start + dynamic discovery”:
- Inject only core tools by default: files, commands, search, messages, task status, and tool discovery.
- Add meta tools such as
tool_search / tool_manifest so the model can retrieve tools as needed.
- Build an index cache for tool schemas and load them by domain/keyword.
- Keep short schemas for high-frequency tools and only summaries for low-frequency tools.
- Once a tool is selected during the session, inject its full schema or execute its adapter.
The skill system already has a similar idea with “index cache + prompt summary limit”: provide a small summary first, expand only when needed. The tool system could follow the same pattern so the base prompt does not grow linearly with the number of tools.
Acceptance Criteria
- The initial prompt for normal chat/status queries becomes significantly shorter.
- Core tools remain usable without an extra discovery step.
- Long-tail tools can be found and executed through tool search.
- Adding more tools does not linearly increase the base prompt size.
Scenario
This is not exactly a bug, and developers can work around it themselves, but it would be better if GA had a native mechanism to control the system prompt/tool prompt size.
As GenericAgent gains more tools, injecting the full tool schema into the prompt on every startup or every request will quickly make the prompt too long, slow down startup, waste tokens, and make the model navigate through many irrelevant tools.
In real use, many tasks only need core tools: file read/write, command execution, file search, messaging, and status queries. Long-tail tools should be discovered and loaded on demand.
Current Pain Points
Suggested Direction
Introduce “core tool warm start + dynamic discovery”:
tool_search/tool_manifestso the model can retrieve tools as needed.The skill system already has a similar idea with “index cache + prompt summary limit”: provide a small summary first, expand only when needed. The tool system could follow the same pattern so the base prompt does not grow linearly with the number of tools.
Acceptance Criteria