◂ tracks TRK-04
learning track
AI agents
Run agents like production software: sandboxing, tool servers, tracing, and guardrails that survive a bad prompt.
0% 7 modules · ~3h
Agent primer reading
An agent loop combines model reasoning with tools, memory, and permissions. Reliability comes from constraining each step, observing what happened, and making recovery paths explicit.
- ▸ Tool schemas are contracts; vague inputs produce fragile tool calls.
- ▸ Sandboxes limit blast radius when generated commands or files are wrong.
- ▸ Tracing turns an agent run into evidence you can inspect, replay, and improve.
Foundations part 01
- 01 ▤ Anatomy of an agent loop reading 15m
- 02 ❯ Linux sandbox lab 20m
- 03 ❯ LLM sampling lab 30m
Tools & serving part 02
- 04 ▤ Tool schemas and MCP reading 20m
- 05 ❯ LLM inference lab 45m
Keeping control part 03
- 06 ⚑ Runaway agent challenge 35m
- 07 ▤ Guardrails and permissions reading 15m