feel very aligned with our vision & Ronak + the awesome Trajectory team’s on practically tackling Continual Learning at scale 🚀
there’s a very good reason why teams are partly building “Observability Shaped” platforms for tackling CL over long time horizons
it’s because Traces are the gold the rich agent action/outcome space that can be mined for high quality signals and incorporated back into agents via:
- Harness Engineering
- Updating Memory/Context Banks for later retrieval
- SFT, Distillation by mining “good traces”
- Building environments to do RL from Traces
A large part of this is aligning with Companies and Users what they actually want their agents to learn over time. And a big part of that is a data mining problem at scales we’ve never seen before very soon as agents get integrated into every piece of work:
- Classifying errors
- Finding good traces
- Capturing product feedback
- Updating user memory
We are in the earliest innings of tackling CL for agents simply because we’re only just now deploying agents into systems that will live with us across month and year timescales
CL practically imo will be a large mix of large scale data understanding + contextually figuring out how/when to to incorporate this new data
Agents experience their world in a way akin to how humans do, the most interesting research questions are how we capture this data and how we use it to selectively update our agents over ultra long time horizons
And it’s fantastic that more great teams are doing that!! :)
