manager agent pipeline

[vitalyshalumov]

How can I implement some logic in inside the manager agent?
For example, an order berween two managed agents, or a human in the loop after final answer of specific agents?

[KeepALifeUS]

Great question! Here are a few patterns for implementing custom logic in manager agents:

1. Agent Ordering (Sequential Pipeline)

You can enforce order by using the manager's system prompt to specify execution sequence:

manager_agent = ManagedAgent(
    agent=CodeAgent(...),
    name="manager",
    description="Orchestrates agents in order: research_agent -> analysis_agent -> writer_agent"
)

Or implement explicit ordering in your workflow by chaining agent outputs:

# Pseudo-code for sequential execution
result1 = agent1.run(task)
result2 = agent2.run(task + f"\nContext from previous step: {result1}")

2. Human-in-the-Loop Pattern

For HITL after specific agents, you can:

def run_with_approval(agent, task):
    result = agent.run(task)
    print(f"Agent output: {result}")
    approval = input("Approve? (y/n): ")
    if approval.lower() != 'y':
        # Re-run or modify
        return run_with_approval(agent, task)
    return result

3. Shared State Pattern (Stigmergy)

Instead of direct agent-to-agent communication, use a shared state that agents read/write to:

shared_state = {"stage": "research", "data": {}}

# Each agent checks state before acting
if shared_state["stage"] == "research":
    result = research_agent.run(task)
    shared_state["data"]["research"] = result
    shared_state["stage"] = "analysis"

This gives you full control over the pipeline flow and makes it easy to insert human checkpoints.

I documented similar patterns for multi-agent orchestration here: https://github.com/KeepALifeUS/autonomous-agents

What's your specific use case? Happy to suggest a more tailored approach!