Skip to main content

Advanced 3 - Discovery mode for large toolsets

Earlier tutorials assume direct tool injection: every tool's full JSON schema is visible up front. That works for small apps. Large toolsets waste context and make the model pick worse tools.

Discovery mode keeps context_builder meta-tools visible and hides most domain tools behind search. The agent finds a tool with search_tools, learns params with get_tool, then calls execute_tool (or calls the tool by name when the runtime routes it).

Catalog meta-tools in this build:

  • search_tools
  • get_tool
  • execute_tool
  • run_parallel
  • background_run
  • use_skill
  • call_app
  • ask_user

There are no list_categories / browse_category tools in the current catalog. Older tutorials that used those names are obsolete.

Force the mode with:

yaml
runtime:
  tool_injection: discovery   # or direct | compact_direct

Otherwise the runtime picks from context budget vs tool schema size. See Tools.

Minimal app sketch​

app.yaml
schema_version: 2
app:
  app_id: discovery-demo
  name: Discovery demo
  version: "0.1.0"

runtime:
  tool_injection: discovery

agents:
  - id: main
    role: worker
    brain:
      provider: ollama
      model: qwen25-7b-gpu:latest
      backend: openai_compat
      config:
        base_url: http://localhost:11434/v1
        api_key: ollama
    system_prompt: |
      Prefer search_tools when you are unsure which tool exists.
      Call get_tool before first use of an unfamiliar tool.

tools:
  modules:
    filesystem: {}
    bash: {}
    web: {}