Astra 智能体最贵的 bug 是「无限乐观」
工具报错后智能体反复重试烧光预算,作者给出三条护栏:最多 20 次工具调用、每个工具最多重试 2 次、重复报错即停。

the most expensive bug in a GPT-6 Astra agent isn’t bad reasoning. it’s infinite optimism. tool fails → agent retries → receives the same error → retries again → burns your entire budget three guardrails fix this: • max_tool_calls: 20 hard-stops runaway execution • retries_per_tool: 2 prevents one broken tool from hijacking the run • stop_on_repeated_error: true ends the loop when nothing is changing autonomous doesn’t mean “let it run forever.” it means knowing exactly when the agent should stop, recover or ask for help. I broke down the full Astra agent architecture below ↓
这条是引用着另一条写的:x.com/mr_kozh/status/2097221341107925350