Skip to main content

Agent Catalog (Overview)

Aragora supports multiple agent backends for vetted decisionmaking. This page summarizes the agent families and how to select them. For the full up-to-date catalog, see AGENTS.md at the repository root or call list_available_agents() at runtime.

Allowlist note: Server-side validation uses ALLOWED_AGENT_TYPES in aragora/config/settings.py. Some registered agents are opt-in and not allowlisted by default.

Agent Families​

CLI Agents (allowlisted)​

Use external CLI tools with local credentials:

  • claude, codex, openai, gemini-cli, grok-cli, qwen-cli, deepseek-cli, kilocode

Direct API Agents​

Call provider APIs directly:

  • Allowlisted: anthropic-api, openai-api, gemini, grok, ollama
  • Opt-in: mistral-api, codestral, lm-studio, kimi-legacy
  • Fine-tuned (opt-in): tinker, tinker-llama, tinker-qwen, tinker-deepseek

OpenRouter Agents (allowlisted)​

OpenRouter-backed unified access:

  • deepseek, deepseek-r1, llama, mistral, qwen, qwen-max, yi, kimi, kimi-thinking, llama4-maverick, llama4-scout, sonar, command-r, jamba, openrouter

External Framework Proxies (allowlisted)​

  • external-framework, openclaw, crewai, autogen, langgraph

Built-In​

  • demo

Choosing Agents​

Use heterogeneous mixes to reduce correlated blind spots:

from aragora.agents import create_agent

agents = [
    create_agent("anthropic-api", name="proposer", role="proposer"),
    create_agent("openai-api", name="critic", role="critic"),
    create_agent("gemini", name="judge", role="synthesizer"),
]

Roles​

  • proposer: Generates initial responses
  • critic: Finds flaws and counterarguments
  • synthesizer: Produces final consensus outputs