Brief in progress — review the script and demo before recording.
This video needs to clearly explain the core mechanics of agents that are crucial to improving how we work with them. We can go into technical detail, because these mechanics matter both when building agents and when working with them.
- Core Concepts: LLM calls run in a loop, producing tool calls (JSON) or a text response that ends the loop. The LLM decides what to do next, while we can programmatically shape the available context, set boundaries, and steer execution.
- Context: what the user provides is not exactly what the model receives. We need to show this through an existing harness or a simple agent with a system prompt, developer messages, tools, context compaction, and prompt caching.
- Sandboxing: show the difference between function calling and programmatic tool calling (code mode), including the benefits and trade-offs.
- Subagents: explain that an agent can itself be a tool, implemented as a nested loop or a separate process, with either a forked conversation or a fresh one. It would also be useful to briefly highlight how difficult it is to build durable agents that can communicate with one another and recover from errors.