Toolsets

AGENT-K uses Pydantic-AI's FunctionToolset for custom tools and relies on built-in tools when a provider supports them. Built-in tools are preferred for web search, memory, and code execution.

Built-in Tools

Use built-in tools for web search, URL fetching, memory, and code execution. AGENT-K exposes helpers for configuring them:

from pydantic_ai import Agent
from agent_k.toolsets import prepare_web_search, prepare_web_fetch

agent = Agent(
    "anthropic:claude-3-haiku-20240307",
    builtin_tools=[prepare_web_search, prepare_web_fetch],
)

Memory (Anthropic only) requires a backend:

from pydantic_ai import Agent
from agent_k.toolsets import create_memory_backend, prepare_memory_tool, register_memory_tool

memory_backend = create_memory_backend()
agent = Agent(
    "anthropic:claude-3-haiku-20240307",
    builtin_tools=[prepare_memory_tool],
)
register_memory_tool(agent, memory_backend)

Code execution is enabled only for supported providers:

from agent_k.toolsets import prepare_code_execution_tool

agent = Agent(
    "anthropic:claude-3-haiku-20240307",
    builtin_tools=[prepare_code_execution_tool],
)

The Evolver also uses MCPServerTool to call Kaggle MCP endpoints for submissions.

Available Toolsets

Kaggle Toolset

Provides Kaggle API operations:

Tool	Description
kaggle_search_competitions	Search for active competitions
kaggle_get_competition	Get competition details
kaggle_get_leaderboard	Get leaderboard entries
kaggle_list_datasets	List competition datasets

Usage:

from pydantic_ai import Agent
from agent_k.adapters.kaggle import KaggleAdapter, KaggleSettings
from agent_k.core.deps import KaggleDeps
from agent_k.toolsets import kaggle_toolset
from agent_k.ui.agui import EventEmitter

config = KaggleSettings(username="you", api_key="your-key")
adapter = KaggleAdapter(config)

deps = KaggleDeps(kaggle_adapter=adapter, event_emitter=EventEmitter())

agent = Agent(
    "anthropic:claude-3-haiku-20240307",
    deps_type=KaggleDeps,
    toolsets=[kaggle_toolset],
)

result = await agent.run("Find featured competitions", deps=deps)

Code Toolset

code_toolset provides helper tools used by the Evolver for code mutation and evaluation workflows. It is composed with the Evolver's own toolset via create_production_toolset.

Combining Toolsets

Use create_production_toolset to apply strict settings and approvals:

from agent_k.toolsets import create_production_toolset, kaggle_toolset

toolset = create_production_toolset(
    [kaggle_toolset],
    prefix="kaggle",
)

Error Handling

Toolsets return serializable data. Errors should return a structured dict or list so the model can respond gracefully.