Agents and tool use
A language model only writes text, but text can name an action. Put a program around the model that carries out the actions it names and feeds back the results, and you have an agent. How that loop works, how tool calls became structured generation, what ReAct and Toolformer showed, how agents are evaluated, and why prompt injection remains unsolved.
50 min
After this lesson you can
- Describe the agent loop (observe, think, act) and write it as a policy over a growing context
- Explain tool calling as structured generation, including grammar-constrained decoding
- Summarise what ReAct and Toolformer contributed, and decode Toolformer's filtering criterion
- Read agent benchmark results (SWE-bench, τ-bench, METR time horizons) and say what they do and don't measure
- Explain direct and indirect prompt injection, and why current defences are not considered solved
In this lesson
- The loop
- Tool calls as structured generation
- What ReAct and Toolformer showed
- Planning and memory
- Measuring agents
- Prompt injection
- What is and isn’t known (September 2026)
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.