메타 에이전트 구축 가이드: 작업 설명만으로 스스로 에이전트를 설계하고 최적화하는 시스템

class AgentConfig(BaseModel):
    agent_name: str = "DesignedAgent"
    objective: str
    planner: PlannerSpec
    memory: MemorySpec
    tools: List[ToolSpec]
    output_style: str = "concise"
    safety_rules: List[str] = Field(default_factory=lambda: [
        "Do not execute arbitrary OS commands.",
        "Refuse harmful/illegal instructions; suggest safe alternatives.",
        "If uncertain, ask for missing inputs or state assumptions.",
    ])

def run(self, task: str, verbose: bool = True) -> str:
    scratch = ""
    self.memory.add("user", task)
    for step in range(1, self.config.planner.max_steps + 1):
        prompt = self._react_step_prompt(task, scratch)
        raw = self.llm.generate(prompt, max_new_tokens=256, temperature=self.config.planner.temperature)
        m = re.search(r"\{.*\}", raw, re.DOTALL)
        raw_json = m.group(0).strip() if m else raw.strip()
        try:
            action = json.loads(raw_json)
        except Exception:
            // ...(중략)
        if action.get("action") == "tool":
            name = action.get("tool_name", "")
            args = action.get("tool_args", {}) or {}
            res = self.tools.call(name, args)
            scratch += f"
[tool:{name}] args={args}
result_ok={res.ok}
result={res.output}
"
        elif action.get("action") == "final":
            return action.get("final", "").strip()

def refine(self, cfg: AgentConfig, eval_report: Dict[str, Any], task: str) -> AgentConfig:
    new_cfg = cfg.model_copy(deep=True)
    if eval_report["flags"]["generic"] or eval_report["flags"]["mentions_max_steps"]:
        new_cfg.planner.max_steps = min(18, new_cfg.planner.max_steps + 6)
        new_cfg.planner.temperature = min(0.35, new_cfg.planner.temperature + 0.05)
    if new_cfg.memory.kind != "retrieval_tfidf":
        new_cfg.memory.kind = "retrieval_tfidf"
        new_cfg.memory.max_items = max(new_cfg.memory.max_items, 200)
    return new_cfg