Continual learning

Watches live agent conversations for silent failures and suggests fixes.

Agnost publishes a free plan and a plain price list for turning an agent's live conversations into fixes and tests, so how much "learning from use" you get at each price can be tried by hand.

Trains small models for one task with reinforcement learning, plus a Python SDK.

Osmosis publishes a free, MIT-licensed Python SDK with a dated changelog every few days, so its claim of retraining a model from its own evaluation results can be read in code and tried on a small task.

Builds tests from an AI agent’s own traces, then tunes the agent against them.

Relai turns an agent's own traces into replayable tests, offers them on a free plan, and publishes before-and-after pass rates on a public benchmark, so its claim that an agent can keep improving without breaking earlier fixes can be checked.