jarviscore

JarvisCore

Python framework for production multi-agent AI. Agents run on a peer-to-peer mesh, remember across sessions, and trace everything they do. Package: jarviscore-framework (import name jarviscore). Python 3.10+.

The one decision that matters: pick the right profile

• CustomAgent: you bring the brain. Deterministic control, you implement the logic. Use for workers with known logic, API-driven services, wrapping existing code.
• AutoAgent: the framework brings the brain. LLM task execution with sandboxed code generation, self-repair, and a verified-work registry. Use when the task is described, not coded.

There is nothing in between. Do not simulate one with the other.

Minimal AutoAgent (correct, complete)

import asyncio
from jarviscore import AutoAgent, Mesh

class ResearcherAgent(AutoAgent):
    role = "researcher"                      # required
    capabilities = ["research", "analysis"]  # required
    system_prompt = "You are a rigorous research analyst."  # required

async def main():
    mesh = Mesh()
    mesh.add(ResearcherAgent)
    await mesh.start()
    result = await mesh.run_task(agent="researcher", task="Compare SWIM and Raft.")
    print(result)

asyncio.run(main())

Production AutoAgents should also declare description, capability_descriptions, and capability_contracts when peers or distributed goals need to route and authorize work. Optional execution attributes include output_schema, goal_oriented = True (classifies complex tasks into Plan, Execute, Evaluate), default_kernel_role = "...", and requires_auth = True (injects Nexus-backed _auth_manager).

Minimal CustomAgent

from jarviscore import CustomAgent

class ProcessorAgent(CustomAgent):
    role = "processor"
    capabilities = ["processing"]

    async def execute_task(self, task):
        return {"status": "success", "output": ...}

    async def on_peer_request(self, msg):          # P2P handler
        return {"status": "success", "result": ...}