Skip to main content

Agent Workflows

Velane Bun and Python executors ship with Mastra (Bun) and LangGraph (Python) pre-installed. Use them directly in workflow code — no Velane wrapper packages.

Python — LangGraph​

Store OPENAI_API_KEY (or other provider keys) as tenant secrets. They are injected as environment variables at invoke time.

import os
from langchain_openai import ChatOpenAI
from langgraph.graph import StateGraph

def handler(input: dict) -> dict:
    model = ChatOpenAI(model="gpt-4o-mini", api_key=os.environ["OPENAI_API_KEY"])
    # Build and run your graph; return structured output.
    return {"message": "agent result"}

For streaming, use an async generator handler and yield chunks (see Invocation Modes).

Bun — Mastra​

import { Agent } from '@mastra/core/agent'

export default async function handler(input: Record<string, unknown>) {
  const agent = new Agent({
    name: 'workflow-agent',
    instructions: 'You are a helpful assistant.',
    model: 'openai/gpt-4o-mini',
  })
  const result = await agent.generate('Summarize: ' + String(input.topic ?? ''))
  return { text: result.text }
}

Wire integration tools manually with @velane/integrations inside Mastra tool execute functions (see Integrations Overview).

Runtime limits​

Each workflow version stores:

FieldDefaultEnforcement
timeout_ms60000 (1 min)Hard kill → timeout
max_memory_mb200Hard kill → oom_killed
max_cpu_percent10 (0.1 vCPU)cgroup CPU cap (Linux executors)

Set limits in the admin portal Settings tab for a workflow, or via POST /v1/snippets/{id}/versions when creating a version.

Tenant caps (GET /v1/tenant/runtime-limits) are read-only in the UI. Platform operators adjust per-tenant caps in the database (tenants.runtime_limits).

Agent workflows that import LangGraph or Mastra typically need higher max_memory_mb (for example 512–1024). Simple workflows that do not import agent libraries keep a low memory footprint even though the frameworks are present in the executor image.