If a language model only predicts text, how is a chat assistant searching the web and reading your inbox. The answer is tools, and Bani Chaudhuri spends this one on how they actually work. A tool is an external capability attached to the model by the system around it. The model keeps doing what it is good at, understanding the request and deciding what is needed. The system around it does the part the model cannot.

She traces one weather question end to end, then builds the same thing from the other side, as someone creating a scheduling agent. Define the tool and describe it, the user asks a question, the model decides a tool is needed and with what parameters, the platform or your own code runs it, the result comes back, and the model turns that result into a normal sentence. The line to keep is that the model is the coordinator and not the executor, which is why a tool with a vague description gets called at the wrong moment or never gets called at all.

What this video covers

• What tools are, and why a model cannot reach the outside world without them
• What happens behind the scenes when a chat assistant answers a question about live information
• The six steps of a tool call in an agent you build yourself, including what happens when the tool fails
• Why the model coordinates while the platform or your code executes
• What a support agent can do with tools connected that it cannot do without them
• Why the tool description matters as much as the tool, and four ways to write better ones

Chapters

• 0:00 How a chat assistant searches the web
• 0:30 What tools actually are
• 3:12 Building your own agent with tools
• 5:24 Why the model coordinates and does not execute
• 7:38 Tips for adding tools to your agents