Agent Matrix

The Core Insight of This Project: Achieving Higher Intelligence via Nested Agents.

Initialize

flowchart LR
    launch --> |"method 1"| MASTER>"python master mind"]
    launch --> |"method 2"| UE2["Launch UE Client"]
    UE2 --> Redis3["Join Redis Network"]
    MASTER --> IF1{"Has UE ?"}
    IF1 -->|"Yes"| UE["Launch UE Client"]
    IF1 -->|"No"| UED["Download UE Client"]
    UED --> UE
    UE --> Redis2["Join Redis Network"]
    UE2 --> MASTER2>"python master mind"]
    MASTER --> Redis["Join Redis Network"]
    MASTER2 --> Redis["Join Redis Network"]
    Redis --> Wait["Wait Further Command"]
    Wait --> Begin(("Begin Tasks"))

Loading

python agent tick

flowchart TD
    PyAgent["Agent Init"] --> A["Wait Wake Up Condition 等待唤醒"]
    -->AAA
    subgraph AAA["Thread 新线程"]
        direction LR
        TT1["New Thread Begin 线程启动"]
        TT1 --> BBB
        TT1 --> TT2["Read UUID"] -->
        TT3["Read Call History"] -->
        TT4["LM Call Python Pre-process"] -->
        TT5["LM Call"] -->
        TT6["LM Call Python Post-process"] -->
        TT7["Redis Com and Final Render"] -->
        TT8["Select Next Agents To Wake Up"] -->
        TT9["Wake Up Other Agents"] -->
        TT10["Exit"]


        TT5 --> Ta
        Ta --> TT5
        subgraph BBB["Thread 新线程2"]
        Ta["Redis Com and Stream Render"]
        end

    end

Loading

unreal agent tick

flowchart TD
    subgraph Tick
        direction LR
        Tick_S["Tick Begin"] 
        --> Tick_A["Hear from Redis"]
        --> Tick_B["Update Status"]
        --> Tick_C["Read Event"]
        --> Tick_D["Do Event"]
    end

Loading

agent university base class

flowchart LR
    Z["X"]
    Z --> A["Basic LM Agent"]
    Z --> B["Internet Agent"]
    Z --> C["RAG Agent"]
    Z --> D["Python Code Exe Agent"]
    Z --> E["Memory Agent"]
    Z --> F["CoT Agent"]
    Z --> G["Screen Capture Agent"]
    Z --> H["Windows Exe Agent"]
    Z --> I["Shell Exe Agent"]
    Z --> J["Realtime Voice Read Agent"]
    Z --> J2["Realtime Speaker Agent"]

Loading

Agent Activation Call Cycle

1. upper stream agent proxy call current agent proxy's ___on_agent_wakeup___

2. [___on_agent_wakeup___ (anywhere)] current agent modify msg (msg = on_agent_wakeup+arguments) and send_to_real_agent

4. [_begin_acquire_command (agent_matrix/agent/agent.py)]: real agent receive msg and execute _handle_command

5. [_handle_command (agent_matrix/agent/agent.py)]: real agent receive on_agent_wakeup and execute wakeup_in_new_thread (in new thread)

    def wakeup_in_new_thread(self, msg):
        # deal with the message from upstream
        msg.num_step += 1
        if msg.level_shift == '↑':
            # Case 1:
            # - This agent must be a parent with at leat one child agent,
            #   and all its children have finished their tasks.
            #   It is time that this parent exam all the work done by its children
            #   and decide what to do next.
            downstream = self.on_children_fin(msg.kwargs, msg)
        else:
            # Case 2:
            # - If this agent is a parent (with at least one child agent),
            #   on_agent_wakeup will be called, and its children will handle more work afterwards.
            # - If this agent has no children,
            #   on_agent_wakeup will be called.
            downstream = self.on_agent_wakeup(msg.kwargs, msg)

        # deliver message to downstream
        # (don't worry, agent proxy will deal with it,
        # e.g. chosing the right downstream agent)
        self.on_agent_fin(downstream, msg)

6. [on_agent_wakeup (agent_matrix/agent/agent.py)]: real agent receive and execute on_agent_wakeup

7. [on_agent_fin (agent_matrix/agent/agent.py)]: terminate agent task, prepare msg delivery, register downstream to msg

    def on_agent_fin(self, downstream, msg):
        msg.src = self.agent_id
        msg.dst = self.proxy_id
        msg.command = "on_agent_fin"
        msg.kwargs = downstream
        # for switch agent, add downstream_override
        if downstream.get("downstream_override", None):
            msg.downstream_override = downstream["downstream_override"]
        # for groupchat agent, add children_select_override
        if downstream.get("children_select_override", None):
            msg.children_select_override = downstream["children_select_override"]
        if downstream.get("call_children_again", None):
            msg.call_children_again = downstream["call_children_again"]
        if downstream.get("dictionary_logger", None) and isinstance(downstream["dictionary_logger"], dict):
            msg.dictionary_logger.update(downstream["dictionary_logger"])
        # keep level shift unchanged
        msg.level_shift = msg.level_shift
        self._send_msg(msg)