Agent 执行循环

Agent 是 Llama Stack 应用的核心。它们将推理、记忆、安全和工具使用结合到连贯的工作流中。Agent 的核心是一个精密的执行循环，它支持多步推理、工具使用和安全检查。

Agent 工作流中的步骤

每个 Agent 回合遵循以下关键步骤

初始安全检查: 用户的输入首先通过配置的安全防护进行筛选
上下文检索:
- 如果启用了 RAG，Agent 可以选择从记忆库中查询相关文档。您可以使用 instructions 字段来引导 Agent。
- 对于新文档，它们首先被插入到记忆库中。
- 检索到的上下文作为工具响应在消息历史中提供给 LLM。
推理循环: Agent 进入其主要执行循环
- LLM 接收用户提示 (包含先前的工具输出)
- LLM 生成响应，可能包含工具调用
- 如果存在工具调用
  - 工具输入经过安全检查
  - 工具被执行 (例如，网络搜索、代码执行)
  - 工具响应反馈给 LLM 进行合成
- 循环持续进行，直到
  - LLM 提供不含工具调用的最终响应
  - 达到最大迭代次数
  - 超出 token 限制
最终安全检查: Agent 的最终响应通过安全防护进行筛选

        sequenceDiagram
    participant U as User
    participant E as Executor
    participant M as Memory Bank
    participant L as LLM
    participant T as Tools
    participant S as Safety Shield

    Note over U,S: Agent Turn Start
    U->>S: 1. Submit Prompt
    activate S
    S->>E: Input Safety Check
    deactivate S

    loop Inference Loop
        E->>L: 2.1 Augment with Context
        L-->>E: 2.2 Response (with/without tool calls)

        alt Has Tool Calls
            E->>S: Check Tool Input
            S->>T: 3.1 Execute Tool
            T-->>E: 3.2 Tool Response
            E->>L: 4.1 Tool Response
            L-->>E: 4.2 Synthesized Response
        end

        opt Stop Conditions
            Note over E: Break if:
            Note over E: - No tool calls
            Note over E: - Max iterations reached
            Note over E: - Token limit exceeded
        end
    end

    E->>S: Output Safety Check
    S->>U: 5. Final Response

此过程中的每个步骤都可以通过配置进行监控和控制。

Agent 执行循环示例

这里有一个示例演示如何监控 Agent 的执行

from llama_stack_client import LlamaStackClient, Agent, AgentEventLogger
from rich.pretty import pprint

# Replace host and port
client = LlamaStackClient(base_url=f"http://{HOST}:{PORT}")

agent = Agent(
    client,
    # Check with `llama-stack-client models list`
    model="Llama3.2-3B-Instruct",
    instructions="You are a helpful assistant",
    # Enable both RAG and tool usage
    tools=[
        {
            "name": "builtin::rag/knowledge_search",
            "args": {"vector_db_ids": ["my_docs"]},
        },
        "builtin::code_interpreter",
    ],
    # Configure safety (optional)
    input_shields=["llama_guard"],
    output_shields=["llama_guard"],
    # Control the inference loop
    max_infer_iters=5,
    sampling_params={
        "strategy": {"type": "top_p", "temperature": 0.7, "top_p": 0.95},
        "max_tokens": 2048,
    },
)
session_id = agent.create_session("monitored_session")

# Stream the agent's execution steps
response = agent.create_turn(
    messages=[{"role": "user", "content": "Analyze this code and run it"}],
    documents=[
        {
            "content": "https://raw.githubusercontent.com/example/code.py",
            "mime_type": "text/plain",
        }
    ],
    session_id=session_id,
)

# Monitor each step of execution
for log in AgentEventLogger().log(response):
    log.print()

# Using non-streaming API, the response contains input, steps, and output.
response = agent.create_turn(
    messages=[{"role": "user", "content": "Analyze this code and run it"}],
    documents=[
        {
            "content": "https://raw.githubusercontent.com/example/code.py",
            "mime_type": "text/plain",
        }
    ],
    session_id=session_id,
)

pprint(f"Input: {response.input_messages}")
pprint(f"Output: {response.output_message.content}")
pprint(f"Steps: {response.steps}")