Introduction

What is agentc?

agentc is a compiler and runtime toolchain for building production-ready LLM agents. You describe your agent in a declarative manifest file: the model it uses, the tools it can call, the skills it knows about, and how it is deployed. agentc compiles that description into a self-contained artifact ready for production.

The output is not a wrapper around a library or a running service. It is a compiled artifact that embeds your configuration, your tools, and the entire runtime. You ship one thing, and it is the application.

Core ideas

Manifests over boilerplate. Agents are declared in agent.acl files using HCL syntax with a small set of extensions. The compiler resolves dependencies, generates code, and produces a deployable artifact. You focus on what your agent does, not on how execution, persistence, or serving are wired together.

Archetypes define the output shape. Every agent targets an archetype. The archetype determines what kind of artifact is produced and what build toolchain is required. The standalone archetype produces a cross-platform binary. Other archetypes target different deployment models. Choosing an archetype is choosing a deployment target.

Durable execution. Agents run on a checkpointed graph. State is persisted at each step of execution, so the agent can survive process crashes, restarts, or infrastructure failures and resume exactly where it left off. This guarantee applies regardless of which graph implementation is in use.

Polyglot tools. Tools can be written in JavaScript, TypeScript, or Python, sourced from any Model Context Protocol server, or configured as A2A delegation targets. The compiler handles bundling and embedding where needed. At runtime, the agent dispatches to the correct execution environment automatically.

Build-time and runtime configuration. The manifest separates what is baked into the artifact at compile time from what is resolved from the environment at startup. Sensitive values like API keys are never hardcoded; they are read from environment variables at runtime.

Who is it for?

agentc is for engineers who need agents in production for end users. Agent harnesses such as OpenClaw, Claude Code, or Codex are built specifically for local use and are not designed to serve end users at scale. If you are building an agent for yourself or your team, those tools may be a better fit. If you need to build an agent for customers, or if you have shipped services and want the same operational characteristics for your AI workloads, such as crash recovery, structured observability, and configurable deployments, agentc is built for you.

Next steps

• Concepts in 5 minutes: the smallest mental model before you build.
• Build your first agent: install the compiler and compile an agent.
• Architecture overview: understand how the pieces fit together.
• Manifest reference: every manifest field documented.