**Agent Thinking Paradigm Shift:** The underlying thought pattern of the agent has been fully transitioned to the ReAct paradigm. The agent now makes dynamic decisions during execution and adjusts subsequent decisions in real-time based on the outcomes of its actions.
**Prevention of Endless Execution:** To prevent the agent from blindly executing actions endlessly when it cannot achieve its goals, a penalty mechanism has been designed. At the start of an action, the agent estimates the number of steps required for the task. If the actual number of steps exceeds the expected number, the agent will abandon the task with a probability of **p** and persist with a probability of **1-p**. If the agent chooses to persist, the probability **p** will be multiplied by a penalty factor **beta** for the next execution. If **beta** is greater than 1, it encourages abandoning the task; if **beta** is less than 1, it encourages persistent attempts.
What's Changed
* Adaptive agent by vortezwohl in https://github.com/vortezwohl/CEO/pull/8
**Full Changelog**: https://github.com/vortezwohl/CEO/compare/0.7.3-preview...0.8.1-preview