This project will develop new benchmarks for real-world language agents, deriving tasks from software engineering and writing research papers. These benchmarks will provide realistic evaluations of the capabilities of language agents, while allowing for the development of advanced techniques and methods. We will then explore (1) the implementation of new cognitive architectures for agents (based on the theoretical CoALA framework, previously developed), (2) the development of new memory mechanisms for agents to handle long-horizon tasks, and (3) novel learning mechanisms for self-improving agents.