A tool is an external capability an AI agent calls to act on the world rather than only generate text. It can be a custom function, script, API call, scanner, database, or MCP server. Tools turn model reasoning into measurable action, while the agent controller decides which tool to run and when. section 1.3.1 section 2.3.1
A tool should have a narrow responsibility and remain unaware of the wider execution plan. It accepts structured input, returns structured output, and does not decide its own scheduling or interpret how downstream components will use its result. This separation lets a controller compose multiple tools into a feedback-driven workflow. section 2.5.3 section 2.5.4 section 2.5.5
Listing 2.2 The tool schema
class Tool:
name: str
description: str
def invoke(self, input: Dict[str, Any]) -> Dict[str, Any]:
"""Run the tool and return structured output."""
raise NotImplementedError
A minimal tool contract names the capability, describes it, and exposes an invoke method that accepts and returns structured data. section 2.5.3
Use a durable interface around each tool:
- Declare a name and description so the capability can be identified and selected. section 2.5.3
- Make invoke accept structured input and return structured output under the shared observations schema. section 2.5.3
- Record outputs as artifacts so data movement and agent behavior can be inspected or replayed. section 2.5 section 2.8
- Place safety controls around action, including gates, sandboxing, logging, and kill switches. section 2.8
URL extraction and summarization
One tool extracts URLs from text with a regular expression and returns a urls field. A second tool receives that artifact, counts the URLs, and returns a short summary. The agent runs the first tool, records its result, passes the artifact to the second tool, then stops. section 2.5 section 2.5.4 section 2.5.5
Tools do not make a system an agent by themselves. An agent combines tools with messages, memory, a controller, and artifacts; the controller converts model predictions into an intentional workflow. Shared schemas make those components interoperable and support inspection after execution. section 2.3.1 section 2.5.8
















