The anatomy of an agent: eight building blocks

However fancy an AI agent looks, underneath it's made of a small, fixed set of parts. Once you can name them, agents stop feeling like magic and start feeling like something you can reason about, review, and change with confidence. On Digitorn every agent is made of the same eight building blocks, each with exactly one job. You assemble them visually in Studio, but it helps to know what they are.

#The eight blocks

Identity. The name, icon, colour, and description that make the agent this agent. It's what shows up in your workspace and, if you publish it, on the Hub. Small, but it's the part a human recognises first.

Lifecycle. How the agent runs. Is it a conversation you chat with, or a background worker that fires on a schedule, a webhook, or an event? How many turns can it take? What wakes it up? Everything about when and how long the agent runs lives here.

The agents themselves. An app can be one agent or a team. Each agent has a role, a brain (the model it thinks with), a system prompt (its instructions), and the set of tools it's allowed to touch. A coordinator that plans and delegates, a specialist that does one thing fast, a writer, a checker: they're all declared here.

Tools. What the agent can actually do: read and write files, search the web, run shell commands, query a knowledge base, talk to Slack or Gmail through a connector. You pick these like checkboxes, and each agent only gets the ones it needs. If a capability isn't here, the agent cannot use it, which is exactly what makes an agent auditable.

Security. The boundaries. The behaviour rules (edit only what's been read, plan before acting, block destructive commands), the sandbox, and which credentials the agent is allowed to resolve. This is the block that turns "the AI helped me" into "the AI can't quietly do something I didn't intend".

Interface. Pure display: the theme, the greeting, the live preview pane, any widgets. None of it changes what the agent does; it changes how you experience it. A website builder shows a live page here, a slide agent shows a deck, a coding agent shows a workspace.

Developer affordances. The optional extras for people who want to go deeper: reusable skills, variables, shared includes. Most agents never need this block. It's there for the power users, and it stays out of the way for everyone else.

Flow. The one that changes everything when it's present. Without it, an agent runs the standard loop: the model decides, calls tools, sees the results, and loops. With a flow, the agent runs a graph you laid out, where steps fire in a declared order and the model is just one kind of step. Freeform agent versus scripted scenography, that's the choice this block encodes, which is why it stands on its own.

#Why the separation matters

The point of splitting an agent into eight clean parts isn't tidiness for its own sake. It's what makes agents reviewable.

When every concern has exactly one home, "what does this agent do?" is answerable at a glance. What can it touch? Look at tools. What's it not allowed to do? Look at security. How does it run? Look at lifecycle. You never have to guess whether a change is cosmetic or load-bearing, because display lives in interface and behaviour lives in the blocks that actually run.

That clarity is why non-engineers can safely edit agents, why reviews take two minutes instead of a meeting, and why you can hand an agent to a teammate three months from now and they can understand it without a walkthrough.

#Where you meet the eight blocks

You don't hand-write any of this. In Studio, each block is a part of the canvas: you name the agent, choose its brain, check the tools it can use, flip on the safety rules, and the workspace assembles the whole thing for you. The builder assistant can do it from a plain-language description, and you refine it live.

The eight blocks are just the mental model underneath, the vocabulary that lets you talk about what an agent is. Once you have it, everything else on the platform reads more clearly.

#Further reading

• Why you build agents in a workspace instead of writing code: Why you shouldn't write your agents in code
• A full agent, part by part: How to build your own Claude Code
• Ten shapes you can build today: 10 apps you can build in Studio