Near-Term Problems for Agents

April 24, 2026

#ai #agents #memory #scaling

Right, straight to it. Here are the biggest problems I see facing AI agent systems for the next 6-12 months. Some of these are problems that already have solutions in progress (mostly half-baked or half-delivered). Others are closer to frontier problems that make take quite a while to crack. I’ve put them roughly in order of difficulty, as far as I can tell.

Agent runtime isolation
Long-term, portable memory
Very large contexts
Online learning (LLM agents are stuck with snapshots and hydrated memory, no online learning)
Product surface (adoption requires these solutions to be bundled and packaged, no thought required)

Admittedly, #5 is not the hardest, and it really comes in batches after each of the other problems is solved.

Everything else (e.g. get the agent to make connections between data, get the agent to create reasonable charts, get the agent to analyze sources effectively, etc.) will probably be solved with more capable models. The labs will take care of the model training. Anyone else can solve the other problems.

I’m fortunate to be at FML Inc, where we’re working on these problems (especially the memory and context ones). In a few months we may all be on to new and different problems as we collectively discover them. It’s a great time to be involved in a rapidly evolving field.