agent

The agent block is where you define the agent itself: its name, graph, model, system prompt, and what it is allowed to do. A manifest contains one agent block.

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

  version     = "1.0.0"
  description = "A helpful assistant."

  prompt = "You are a helpful assistant. Your name is {{ agent_name }}."

  capabilities      = ["math", "filesystem"]
  capability_policy = "locked"

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

Fields

graph block

The graph block selects which graph implementation drives the agent. Graph selection is required and currently the only shipped value is react.

graph {
  type = "react"
}

See graph for the block reference and ReAct for the shipped graph's behavior.

prompt

The prompt field sets the agent's prompt. It accepts a plain string, a list of role/content objects for multi-turn prompts, or a configured prompt source.

# Plain string
prompt = "You are a helpful assistant."

# Multi-turn
prompt = [
  { role = "system", content = "You are a helpful assistant." },
  { role = "user",   content = "Always be concise." }
]

The prompt supports Jinja2 template syntax. Variables are injected at runtime before each run. The set of available variables is graph-dependent; see ReAct prompt templates for the list the ReAct graph provides, and Write templated prompts for how to use them.

Langfuse prompt source

Use a Langfuse source object to retrieve the prompt from Langfuse Prompt Management at runtime:

prompt = {
  source = "langfuse"

  prompt_name = "support/assistant"
  public_key  = runtime("LANGFUSE_PUBLIC_KEY")
  secret_key  = secret(runtime("LANGFUSE_SECRET_KEY"))

  label                 = runtime("LANGFUSE_PROMPT_LABEL", "staging")
  cache_ttl_seconds     = runtime("LANGFUSE_CACHE_TTL_SECONDS", 60)
  fetch_timeout_seconds = runtime("LANGFUSE_FETCH_TIMEOUT_SECONDS", 5)
  max_retries           = runtime("LANGFUSE_MAX_RETRIES", 2)
}

label and version are mutually exclusive. When both are omitted, Langfuse applies its default production selection.

See Manage prompts with Langfuse for credential setup, selection strategies, caching behavior, and prompt conversion.

capabilities and capability_policy

Capabilities are tags that control which tools the agent can invoke at runtime. Each tool can declare a set of capability tags. The agent's capabilities list declares which tags the agent is allowed to use.

When capability_policy is set to "locked", the agent can only invoke tools whose capability tags are a subset of its declared capabilities. Tools with no declared capabilities are always available.

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

  capabilities      = ["math", "filesystem::read"]
  capability_policy = "locked"
  ...
}

See Tools and capabilities for how capabilities work with tools and Control tool access with capabilities for the recipe.

model block

The model block selects the LLM the agent uses.

Inference parameters such as temperature, max_tokens, and top_p are declared in the providers block, not here. Provider-level defaults apply to every model request, and individual models can override them with their own params block. See Configure a model provider and the providers reference for details.