Agents¶
简体中文 · English
Follow the path from the next action to a complete execution workflow: planning, memory, tool collaboration, self-improvement, training, and independent evaluation.
Choose a subtopic and start with its paper questions.
Planning and search: what should happen next?
With long tasks and delayed feedback, how should an agent choose its next action and undo an earlier mistake?
Memory and skills: retrieval, reuse, and invalidation
When history exceeds the context window, what should be saved, when should it be retrieved, and how can stale knowledge be kept from misleading a new task?
Tools and collaboration: turning plans into inspectable actions
With a plan in hand, how does an agent choose tools, interpret results, and share reliable execution state across roles?
Harnesses and self-improvement: which layer changes?
After a failure, should the system revise one prompt, save a lesson, or change the entire task workflow?
Interactive policy learning: converting trajectory feedback into updates
If a long task gives only a terminal success or failure, which turn should receive credit?
Evaluation and reliability: how do we know it worked?
How can we distinguish model success, a permissive environment, scorer bias, and unsafe overreach?