What's this about?
Two papers at 8 upvotes each. E-Commerce Bench evaluates agents on long-horizon autonomous business operation — running a shop rather than answering a question about one. Agents in the Large proposes a perception-centred architecture for persistent agents.
What does "perception-centred" mean for a persistent agent?
That the primary design problem is not reasoning but noticing. An agent that runs indefinitely faces a world that changes without telling it: inventory moves, prices change, a service goes down, a customer replies. Most agent architectures are action-centred — decide, act, observe the result. A perception-centred one is organised around continuously ingesting state changes and deciding which ones matter, which is the harder half and the one that determines whether the agent is still connected to reality after a week.
Why does this matter to you?
- Running a business end to end is a much better long-horizon benchmark than another coding task, because the failure modes are financial and legible.
- Persistence is the property that turns an agent into infrastructure, with all the monitoring obligations that implies.
- Both are early and lightly upvoted; noted as an emerging direction rather than a result.