What is Blobfish.AI?
Blobfish AI is an RL environment factory designed for frontier labs that need to train agents to reliably operate enterprise software such as CRMs, ERPs, legal systems, and procurement tools. The core challenge in agentic RL is not the algorithm but the supply of realistic, executable environments. Blobfish solves this by generating runnable gyms with executable rewards directly from production evidence.
With a single command, Blobfish packages production traces, tool definitions, database schemas, and skill instructions into a mirror bundle. The mirror builder compiles this snapshot into an executable world that runs as a standard MCP server. Every trace is replayed to prove parity, ensuring the environment faithfully replicates real enterprise complexity—including seasonal pricing, loyalty programs, partial refunds, multi-currency, legacy data formats, and conflicting policies. This eliminates the gap between clean mocks and messy production.
Key features include: no root installation (Python 3.12+ and git required), CLI and MCP server installed to ~/.blobfish, and integration with Claude Code or any agent. The environment provides binary, executable rewards—did the right state change happen in the right system?—avoiding noisy LLM-as-judge signals. This enables a tight RL loop grounded in real outcomes.
Use cases span from training agents on enterprise workflows to adding regression gates in PR pipelines. If a skill or tool change breaks production trace parity, the build fails before reaching customers. Frontier labs can scale from hand-built environments to thousands, as research shows scaling from 100 to 2,000 environments yields a +20.1 point lift on agentic benchmarks. Blobfish is a factory, not a services team, turning real evidence into gated worlds for training. Imported mirrors remain replay/evaluation artifacts until regenerated through the deep path, supporting iterative refinement.
Technical details: The world runs as a standard MCP server, registered in any agent framework. The snapshot includes tool definitions, database schema, production traces, and skill instructions. The mirror builder ensures every trace replays correctly, proving parity. This approach supports continuous integration and automated regression testing, making it ideal for research teams and production deployments alike.
Use Cases
Key Features
RL environment factory production trace replay executable reward generation mirror bundle creation MCP server integration enterprise software simulation multi currency support legacy data handling policy conflict resolution no root installationWho is it for?
frontier AI labs, RL researchers, enterprise software trainers, CRM/ERP developers, legal system engineers, procurement tool teams, agent training specialists
