Agents and prompts

An agent is the center of a manifest. Everything else exists to serve it: the providers give it a model to think with, the tools give it actions to take, the skills give it judgement, and the graph gives it a loop to run in. This page explains how the pieces of the agent itself fit together, so the guides that configure each one have a shared frame.

Identity

The agent block names the agent and describes it. The name is more than a label: it is available to the system prompt at runtime and is used as the service name in traces. The version and description travel with the artifact and appear in diagnostics. A manifest defines a single agent.

The prompt

The system prompt is the agent's standing instructions. In agentc the prompt is not a static string baked in and forgotten; it is a template rendered before every run. That lets one prompt adapt to the agent's identity, the skills currently registered, per-request context sent by the client, and the current time.

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

  prompt = <<-EOT
    You are {{ agent_name }}, a helpful assistant.
    {%- if context_vars %}
    Some context about who you are helping:
    {%- for var in context_vars %}
    - {{ var.description }} is {{ var.value }}
    {%- endfor %}
    {%- endif %}
  EOT
}

Here {{ agent_name }} and the context_vars loop are filled in at runtime. Context variables are values the client attaches to a run, so the same prompt can greet a signed-in customer by name or note their account tier without any change to the manifest. A client might send a context variable whose description is "the user's name" and whose value is "Ada", and the loop above renders it into the prompt. You can also provide several messages instead of one, seeding the conversation with more than a single system instruction.

The set of variables you can reference is not fixed by agentc as a whole; it is defined by the graph the agent runs on, because the graph decides what state exists during a run. The variables available to the ReAct graph, such as agent_name, skills, context_vars, and current_datetime, are listed in the ReAct prompt templates reference. The mechanics of writing these templates are covered in Write templated prompts, and sending per-request values from a client is covered in Pass context from the client.

The model

The agent selects a model by naming a provider and a model identifier. The provider must be one declared in the providers block, which is also where inference parameters such as temperature and token limits live. Separating the provider configuration from the agent's model selection means the same agent can point at different models by changing one field. See Configure a model provider.

Capabilities

The agent declares which capabilities it holds, and those capabilities decide which tools it may invoke. This is the boundary between what the agent could call and what it is allowed to call. The concept is covered in Tools and capabilities, and the setup recipe in Control tool access with capabilities.

Where to go next

• Write templated prompts: make the system prompt adapt at runtime.
• Configure a model provider: choose a model and set inference parameters.
• agent reference: every field of the agent block.