Concepts in 5 minutes

Before you build anything, it helps to hold a small map of how agentc works. This page gives you just enough vocabulary to follow the rest of the Get started guides. You do not need the full picture yet; the Concepts section covers everything in depth.

agentc compiles a description into a program

You describe an agent in a single file, and agentc turns that description into a program you can run. There are three things to name:

• The manifest is a file named agent.acl. It describes your agent and is the only thing you write by hand.
• The compiler is the agentc command. It reads the manifest and produces the artifact.
• The artifact is the output. For the default setup it is a single self-contained binary.

You write the manifest, run agentc build, and get a binary. That binary is the whole application.

What is inside an agent

An agent is made of a few parts, all declared in the manifest:

• A graph, which is the reasoning loop the agent runs in. Today the graph is react, which calls the model, lets it use tools, and repeats until the task is done. Every graph is checkpointed, so a run survives a crash and resumes where it left off.
• A model, provided by a provider such as Anthropic. The provider is where credentials and inference settings live.
• Optional tools, which are actions the agent can take, such as calling an API or running a command.
• Optional skills, which are Markdown documents that guide how the agent behaves.

A minimal manifest brings the first two together:

build {
  archetype = "standalone"
}

providers {
  anthropic {
    models = ["claude-haiku-4-5"]
    config {
      api_key = secret(runtime("ANTHROPIC_API_KEY"))
    }
  }
}

agent "my_agent" {
  graph {
    type = "react"
  }

  prompt = "You are a helpful assistant."

  model {
    provider = "anthropic"
    name     = "claude-haiku-4-5"
  }
}

Build-time and runtime

Notice runtime("ANTHROPIC_API_KEY") above. Some values are fixed into the artifact when you compile, and others are read from environment variables when the binary starts. A runtime() value is read at startup, and wrapping it in secret() keeps it out of logs and diagnostics. This is what lets you build one binary and run it in different environments by changing environment variables, never recompiling.

Two ways to run the artifact

The compiled binary has a small command-line interface with two commands you will use next:

• run executes a single request and prints the answer, then exits. It is the fastest way to see your agent work.
• serve starts the built-in HTTP server so clients can talk to the agent over HTTP.

What you will do next

In Build your first agent you will install agentc, scaffold a project, compile it, and run your agent for the first time.