Accept prebuilt agent objects and let agents declare dependencies

[nickhuo]

Draft. Two changes that let a LangChain / LangGraph / CrewAI project run on Ventis without being rewritten.

1. _load_agent no longer requires a class

agent_class = getattr(module, self.agent_name)
agent_instance = agent_class()

This requires agent.name to be a class with a no-arg constructor. What the frameworks actually export is an object:

Framework	Exported	Type
LangGraph	graph = builder.compile()	CompiledStateGraph
LangChain LCEL	chain = prompt | llm | parser	RunnableSequence
CrewAI	crew = Crew(agents=..., tasks=...)	Crew

Calling one raises TypeError, _load_agent swallows it, and every request answers "No agent loaded". The only way through was an adapter class that exists purely to hold a reference and forward one call.

Now:

target = getattr(module, self.agent_name)
agent_instance = target() if isinstance(target, type) else target

Two consequences:

• The adapter disappears. A yaml can name the exported graph directly and the port needs no Python for the agent at all.
• Configuration gets somewhere to live. agent_class() takes no arguments, so a configured agent previously had nowhere to receive its config except environment variables read inside __init__. A project can now export agent = MyAgent(model="gpt-4.1") and Ventis uses it as-is.

Backward compatible: isinstance(target, type) is true for classes, so all four existing examples take the same path as before.

2. Agents and workflows can declare dependencies

generate_docker and generate_workflow_docker each wrote a fixed requirements.txt (hard code)

  - name: email_assistant
    entrypoint: src/email_assistant.py
    requirements:
      - langchain>=1.0.0
      - langgraph>=1.0.0
      - python-dotenv

Declared entries are appended to the runtime's own list. An entry naming a runtime dependency is dropped rather than appended, so a project cannot repin something the controller has to run against (redis>=5.0 is ignored, not added).

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}