Module 14 of 16
Agentic systems
An agent is a model choosing its own next step in a loop. This module covers what the system around that loop must guarantee: bounds, memory, durability, and control over what the agent may do.
Once the model picks the next step, a request becomes a long-running, stateful, variable-cost process that can change things in the world. The model supplies judgment; the harness around it decides how long it runs, what it remembers, whether it survives a crash, and which actions it may take.
The four lessons cover the loop and how it stops, memory and planning and what they cost, durable execution for tasks that outlive a process, and the policy and approval layer between a proposed action and its effect. Afterwards you can size an agent’s cost distribution, choose a memory design and name its failure modes, make a task resumable without repeating side effects, and bind a human approval to the exact action it authorizes.
Lessons
- 14.aThe agent loopWhat is the agent loop, and how do you bound it, stop it, and recover it when steps fail?19 min
- 14.bMemory and planningHow does an agent remember and plan over a long task, and what does each design cost and break?18 min
- 14.cDurable executionHow does a long agent task survive crashes, deploys, and day-long pauses without repeating its side effects?16 min
- 14.dTools, permissions, and approvalHow do you decide what an agent may do, and how do you make a human approval mean what the human saw?16 min