Design an Agentic Tool-Use System

This shows up on the infra, ML-systems, and applied-AI tracks, usually phrased as an agent that "autonomously adapts to new tasks." The shape underneath is a loop. A frozen model receives a goal plus a list of tools it is allowed to use. It answers with either a final response or a request to call one of those tools. The system runs the tool and feeds the result back to the model, and the model decides what to do next. Chat answers in one turn. An agent runs that turn over and over, accumulating results, until it judges the goal met or the system cuts it off. Everything in the design serves that loop. How does the model see its tools, how does a tool run safely, how does the growing transcript stay inside the token budget, and how does an hour-long task survive a crash?

The Anthropic angle

The "adapts to new tasks" clause has a concrete answer at Anthropic: you add a server, you do not touch the agent. Anthropic authored the Model Context Protocol, an open standard for exposing tools and data to a model. The architecture is host, client, server: the orchestrator is the host, it holds one client per connection, and each tool or data source runs behind an MCP server that speaks the same protocol. The agent talks to a stable interface, so giving it a new capability means standing up another server rather than editing the loop. Name MCP and say plainly what it decouples, the model from N bespoke integrations, and you are designing on the protocol Anthropic shipped. Anthropic's engineering post Code execution with MCP names the cost: most MCP clients load every tool definition into context up front, and every intermediate result passes back through the model.