Description
Blobfish AI describes itself as a training gym for AI agents. Instead of shipping a product for end users, it builds simulated enterprise environments where AI models can practice real workflows, such as investigating a spreadsheet discrepancy or acting inside a mock CRM or ERP, then receive exact feedback on whether they completed the task correctly.
Read More
Each environment includes realistic mock data (connected records, files and state that behave like a real workplace), tools the agent can call through mock MCP servers with a uniform CLI, training tasks with adjustable difficulty, and deterministic verifiers that check whether the agent changed the simulated world correctly, without relying on another AI model as a judge.
The company builds environments by first proving a tool workflow works, mapping how the available tools can be combined, sampling a valid sequence of tool calls with a known outcome, then writing a realistic task request backward from that known path, so the correct answer and its verifier exist from the start. This is meant to address what the founder describes as the core bottleneck in agent post-training: useful tasks become too easy for a model after a training round, forcing teams to spend months rebuilding tool environments and curricula from scratch. The founder's own origin story, drawn from six months spent building enterprise post-training environments for a financial-services client, describes mapping around 30 expert workflows and finding only about six were genuinely useful for reinforcement learning, sitting at the model's capability boundary: hard enough to expose failure, possible enough to reward progress. Blobfish frames its offering as a continuous loop that regenerates new, appropriately difficult tasks as a model's capabilities improve.
The company also publishes benchmarks, a public leaderboard, and datasets, including MCP-specific benchmarks, alongside its core simulation product. Blobfish AI pivoted from an earlier product, an AI-powered outbound sales dialer; the current site and offering are entirely focused on agent training infrastructure for enterprise workflows, not sales automation, and are aimed at AI labs, model developers and enterprise teams rather than a general business audience.
- CategoryMLOps & Model Deployment
- DeploymentCloud
- PlatformsWeb
Blobfish AI (blobfish.ai)'s Core Features
Simulated enterprise environments (CRMs, ERPs, spreadsheets) for AI agent training
Mock MCP servers with a uniform agent CLI
Realistic mock data: connected records, files and workplace-like state
Adjustable-difficulty training tasks mapped to a model's capability boundary
Deterministic verifiers that check task outcomes without an AI judge
Continuously regenerated tasks as model capabilities improve
Public benchmarks, leaderboard, and datasets, including MCP-specific benchmarks
How to use Blobfish AI (blobfish.ai)?
Talk to the founder: Contact Blobfish AI directly to scope a training environment for your model.
Define the workflow: Work with Blobfish to map an enterprise workflow (CRM, ERP or spreadsheet task) worth training on.
Generate the environment: Blobfish builds the mock tools, data, and a valid task path for the workflow.
Train the model: Run your AI agent against the generated tasks and verifiers.
Check the leaderboard: Compare results against Blobfish's public benchmarks and datasets.
Blobfish AI (blobfish.ai)'s Use Cases
- Agent reinforcement learning
- Post-training curriculum refresh
- Agent capability benchmarking
