Roadmap

What we release today is Phase 1: the autoresearch loop. These are the pieces we're building next to turn it into the full continual-learning system.

Continual-learning autoresearch abstractions

The autoresearch loop is the engine; the next layer is the abstractions that make continual learning autonomous. As laid out in our whitepaper, the end state is a closed loop: a deployed model produces production traces; a lightweight decision agent judges whether a batch is worth learning from and, if so, triggers autoresearch - a heavier, budgeted search for the right training recipe.

Because there is no single algorithm that enables continual learning across all data - different data demands different recipes - the autoresearch agent must autonomously run experiments, recover from training collapses, and decide which checkpoint is finally worth deploying, gated by deployment evals before anything ships. We're building these abstractions directly on top of the SDK, so continual learning graduates from a config modifier into a self-driving loop.

--remote - run the whole client in a sandbox

Today you run the SDK client locally and it drives the training backend over the network. The --remote feature will run the entire client side inside an e2b sandbox - so nothing runs on your machine. No local Python environment, no local setup: launch an experiment and the whole thing executes remotely, for maximum ease of use.

Launch jobs on your own GPUs

A job orchestrator so you can launch training jobs on your own GPUs, not only a hosted Tinker-compatible service. Bring your own compute and run the same configs against it - the SDK handles scheduling and bookkeeping, the GPUs are yours.