Blobfish AI

Blobfish AI addresses a specific bottleneck in agent post-training: building the first task set is manageable, but rebuilding the next one after the model improves is not. Once a model clears a curriculum, those tasks stop teaching anything, and gathering fresh ones through specialist interviews and hand-built sandboxes can take months. Blobfish replaces that manual reset with a continuous generation loop.

Each environment is a functioning world rather than a static dataset. It contains connected records, files and state that behave like a real workplace, mock MCP servers with a uniform agent CLI covering CRMs, ERPs and spreadsheets, grounded workflows with adjustable difficulty and known outcomes, and deterministic verifiers that check whether the agent changed the world correctly. Because the expected end state is generated before the task itself, verification does not depend on an AI judge.

The generation method is 'generate forward, author backward': build the simulated world, map a graph of valid tool interactions, sample a viable path of tool calls with a known final state, then write the user request that would naturally require that path. The company reports early experiments generating 2,000 domain-specific environments to train a 14-billion-parameter model that outperformed a much larger DeepSeek model on MCP benchmarks. It is aimed at post-training and research teams rather than end users, so engagement starts with a conversation rather than self-serve signup.