independent research project
search-guided agents for multi-errand planning under tool failures
status: in progress
- problem
- when an ai assistant's tools fail (a lookup times out, a response comes back incomplete), can it still plan a workable route for several errands?
- approach
- compares an agent that plans with a* search and replans after errors against a react-style agent that calls tools step by step. everything runs offline against a frozen map and simulated tools, so experiments repeat.
- measured
- success rate, extra tool calls, and optimality gap under injected failure rates of 0%, 15%, and 30%.
- outcome
- experiments are still being set up. results will be posted here.