AI Explained
From one answer to an agent
An agent is a model that chooses its own next step. Powerful, and not always what you need.
The loop
What makes an agent an agent
An agent is a model that chooses its own next step. Give it tools and a task, and it works in a loop until the task is done.
Press Play to walk through it, or Listen to have it read to you.
Read all the steps
- A person gives the task once.
- The model chooses a tool and calls it.
- The tool does something: it searches, reads a file, runs a command.
- The model reads what came back and decides what to do next. This is the loop.
- It goes round as many times as the task needs. Nobody wrote the steps ahead of time.
- It stops when the task is finished, when it needs a person's decision, or when it hits a limit.
Workflow or agent
In a workflow, models and tools follow paths written in code ahead of time. In an agent, the model directs its own steps and its own use of tools [1].
Start simple
The advice from the people who build them: find the simplest solution that works, and add a multi-step system only when the simple one falls short [1].
The cost of autonomy
Agents cost more and their errors can compound, so the same source recommends testing them in a sandbox with guardrails [1].
Long tasks
Managing what the agent sees
On a long task the window fills with old results. What to keep, what to summarize and what to move out is most of the engineering.
Context engineering
The strategies for choosing and maintaining the best set of tokens in the model's window while it works: everything beyond the prompt itself [2].
An attention budget
As the window fills, the model recalls what is in it less accurately. Context is a limited resource, and every added token spends some of it [2].
Compaction
When a conversation nears the limit, summarize it and start a fresh window with the summary [2].
Notes outside the window
The agent writes notes somewhere that persists and reads them back when it needs them [2].
Subagents
Specialized subagents work in clean windows of their own and return short summaries [2].
Go deeper
Seven shapes, in 3D
The seven shapes an AI system can takeFrom one model call to a full agent: turn each one, play a request through it and see where an attack gets in. On the AI Delivery side of this site.Spec-driven developmentWrite the spec first, then let the agent build. With a spec builder.Sources
Where this comes from
Read 2026-11-15, unless a source gives its own date. A number in brackets points to the page a statement was read on. Sentences with no number, the small demonstrations and the tools are this center's own explanations.