Tool use
How agents select and use tools without losing the goal.
Research
Small experiments on useful and reliable AI.
We explore practical questions about AI agents and document what works, what fails and what remains uncertain.
Research
01
Current areas
Planning
How longer tasks can be divided into visible, checkable steps.
Generalisation
Whether a method still works when the task changes.
Evaluation
Methods for measuring whether a system is useful and reliable.
02
Questions we care about
- 01Can an agent explain why it chose a tool?
- 02Where should a person review the work?
- 03Which failures repeat across different tasks?
- 04Can small tests predict real-world usefulness?
Public research notes will state the method, evidence, current status and remaining uncertainty.