Blobfish AI (blobfish.ai)

Blobfish AI builds simulation infrastructure for training AI agents. Inside generated environments, AI models practice real enterprise workflows such as CRM, ERP, and spreadsheet tasks: they attempt tasks, fail, receive feedback, and improve over time. It is aimed at teams that train their own models and run reinforcement learning, rather than casual end users.

Blobfish first builds a simulated world with realistic tools, data, and relationships, then constructs a graph of how those tools can interact. It samples a valid sequence of tool calls with a known final state, writes a realistic request an employee might delegate, and checks the result with a deterministic verifier rather than an AI judge.

Each environment ships with realistic mock data, mock MCP servers behind a uniform agent CLI, training tasks with adjustable difficulty, and deterministic verifiers that check whether an agent changed the world correctly. The API exposes 91 endpoints across 14 categories to generate, run, and verify these synthetic RL worlds over HTTP, and a Claude Code skill can generate and serve local practice worlds offline.

Read endpoints work without an API key, and generating worlds is free with a signed-in account; a free key adds scoped world ownership. Quality checks gate training-readiness, blocking benchmarks and queue-for-training until degenerate tasks and missing research coverage are fixed. A server-side watchdog fails any job with no progress for eight minutes.

Use cases

• Training a model with reinforcement learning on enterprise workflows
• Evaluating a production agent against mock enterprise tools before deployment
• Scoring concurrent agent rollouts in a Langfuse dataset run
• Practicing agent workflows locally without a hosted job