Configuration Reference

Agents are defined by an agent.yaml file in your project root. This page documents every available field.

Full example

name: support-agent
version: 0.1.0

models:
  primary: claude-sonnet-4-5
  fallback:
    - gpt-4o
  temperature: 0.3
  max_tokens: 4096

tools:
  - name: create-ticket
    type: http
    endpoint: https://api.linear.app/issues
    auth: env:LINEAR_API_KEY

prompts:
  system: ./prompts/system.md

evals:
  - name: helpfulness
    type: llm-judge
    model: claude-sonnet-4-5
    criteria: "Is the response helpful and relevant?"

Fields

name

Required. The agent’s unique identifier within your team. Must be lowercase alphanumeric with hyphens.

name: support-agent

version

Required. Semantic version string for the agent configuration.

version: 0.1.0

models

Required. Model configuration for the agent.

Field	Type	Description
primary	string	Primary model to use (e.g., claude-sonnet-4-5, gpt-4o)
fallback	string[]	Ordered list of fallback models
temperature	number	Sampling temperature (0-2)
max_tokens	number	Maximum tokens in the response

tools

Optional list of tools the agent can invoke.

Field	Type	Description
name	string	Tool identifier
type	string	Tool type: http, function
endpoint	string	URL for HTTP tools
auth	string	Auth reference (e.g., env:API_KEY)

Tool auth values prefixed with env: are resolved from secrets stored in the AgentLift platform. Set secrets via the dashboard or API.

prompts

System prompt configuration.

Field	Type	Description
system	string	Path to the system prompt file, relative to project root

The system prompt file supports Mustache-style template variables that are resolved at runtime.

evals

Optional list of evaluation criteria.

Field	Type	Description
name	string	Eval identifier
type	string	Eval type: llm-judge, exact-match, custom
model	string	Model to use for LLM-judge evals
criteria	string	Natural language criteria for LLM-judge evals