JackHopkins
Time-constrained tasks
One of the most important trade-offs for real-world engineering is that of time spent in planning vs doing. FLE is an excellent environment to test agents' ability to effectively weigh thinking/planning more (exploration) vs. taking the best actions they've already thought of (exploitation).
In practice this could be extensions to UnboundedThroughputTask such as "produce as many iron plates as possible within 5 minutes", where there is a real-time measurement applied on the environment steps. This could be a combination of either/both wall clock time or ticks of in-game time, but something that implies a cost for the planning and/or actions of the agent.
It would be interesting to see how agents allocate their test-time compute when there are explicit tradeoffs about doing so. Time-constrained activity is more realistic for concurrent multi-agent settings as well, where we'll want to see if multiple agents coordinating together actually save time or other resources as a result.
