Manual today
3–6 monthsExperts describe the work. Engineers recreate it. Researchers find the few useful tasks.
Simulation infrastructure for agent training
A training gym for AI agents.
AI models practice real enterprise workflows inside simulated tools, learn from exact feedback, and keep training as their capabilities improve.
Built from first-hand experience.Six months on xAI’s Grok Enterprise Post-Training team.
01
The bottleneck
The hard part is not building the first task set. It is building the next one.
After post-training, useful tasks become too easy. Teams then spend months interviewing specialists, rebuilding tool environments, and validating a new curriculum. Blobfish turns that reset into a continuous loop.
The Blobfish loop
Continuously generatedWorlds, workflows, difficulty levels, and exact verifiers evolve with the model.
02
What we deliver
A working world, not a static dataset.
Each environment is ready for an agent to enter, act, fail, and try again.
Realistic mock data
Connected records, files, and state that behave like a real workplace.
Tools agents can use
Mock MCP servers, CLI tools, CRMs, ERPs, and spreadsheets.
Training tasks
Grounded workflows with adjustable difficulty and known outcomes.
Deterministic verifiers
Exact checks for whether the agent changed the world correctly.
From world to verified task
Generate forward. Author backward.
We first prove that a tool workflow works. Then we turn that known path into a realistic request, so the answer and verifier exist from the start.
- 01Build the world
Create realistic tools, data, permissions, and relationships.
- 02Map interactions
Construct a graph of valid ways those tools can work together.
- 03Generate a path
Sample a valid sequence of tool calls with a known final state.
- 04Write the request
Work backward into a realistic task an employee might delegate.
- 05Verify the result
Check the expected state change without relying on an AI judge.
03
Founder Q&A
Why this company exists
The problem, in the founder’s own words.
The short version is above. The complete answers are here for customers and researchers who want the operating detail behind the product.
01
What are you building?
We’re building a training gym for AI agents.
Inside our simulated environments, AI models can practice real enterprise workflows using tools like CRMs, ERPs, and spreadsheets. They can attempt tasks, fail, receive feedback, and improve over time.
30 → 6
Only six of 30 mapped workflows were near the model’s capability boundary.
One round later
Those useful tasks often became too easy, forcing the whole cycle to restart.
Read the complete origin story
The idea came from my work at xAI. I left five weeks ago, and before that, I spent six months on the Grok Enterprise Post-Training team.
One project involved helping Morgan Stanley train Grok on complex financial workflows. For example, we might give the model a balance sheet in which the final number is off by 3% because of a currency-conversion issue. The model then has to investigate the spreadsheet and figure out exactly what went wrong.
This isn’t a one-shot question. Solving it requires multiple steps across different files and tools. Like any machine-learning system, when the model encounters workflows outside its training distribution, its performance drops significantly.
To train the model on these workflows, we first had to interview accountants at Morgan Stanley and identify around 30 tasks they perform regularly. Then we manually created Docker-based sandbox environments containing mock spreadsheets, mock Salesforce data, and all the tools the model needed to complete those tasks.
But out of those 30 tasks, only six were useful for reinforcement learning. Some were too easy, so the model always passed. Others were too difficult, so the model always failed. Only six were near the model’s capability boundary, where reinforcement learning could actually help.
After one round of post-training, those six tasks often became too easy. We then had to create another set of challenging tasks and repeat the entire process.
We encountered the same problem when building workflows for companies like PayPal and law firms such as Gibson Dunn.
At xAI, I proposed automatically generating these environments and continuously testing models to find tasks at the edge of their capabilities. The project wasn’t funded because maintaining thousands of MCP integrations, CLI tools, and simulated versions of constantly changing enterprise software would require a dedicated team.
But my manager strongly believed in the idea, so I decided to leave and build it independently.
We’re now getting feedback from people who worked directly on this problem. Nikunj from the Grok Enterprise Post-Training team is reviewing our iterations, and Boyan, a former colleague who joined Reflection AI, is helping validate our generated data. Mercor has also approached us about using our API to accelerate its simulation-creation process.
02
How do Mercor and other companies create these environments today?
The process is still extremely manual.
Mercor has a network of around 50,000 specialists. It asks these specialists to describe approximately 10 tasks they perform in their jobs every day. Engineers then translate those descriptions into simulated environments containing mock data, tools, tasks, and verifiers.
That requires both domain experts and engineers. It is expensive, difficult to scale, and can take months before an environment is ready for training.
03
What exactly would you sell to Mercor or xAI?
We provide ready-to-use, domain-specific simulation environments for training AI agents.
- Realistic mock data
- Mock MCP servers and CLI tools
- Training tasks
- Deterministic verifiers
- Different levels of task difficulty
Our key insight is that we can create high-quality environments without requiring a domain expert to manually design every task.
See how generation works
We begin by constructing a simulated world with realistic tools, data, and relationships. We then build a tool-correlation graph that captures how those tools can interact.
From that graph, we generate a valid sequence of tool calls. Because we already know the correct sequence, we can work backward and generate a realistic user request that requires the model to follow that workflow.
We then apply a realism gate: Does this request make sense? Is it something a real employee might ask an AI agent to do?
Because the expected state changes are already known, the verifier is deterministic. We can automatically determine whether the model completed the task correctly.
This lets us generate large numbers of realistic, verifiable training tasks much faster than the current expert-driven process.
04
If a development shop charges $40,000 to $50,000 per environment, why not sell the environments manually first?
We can—and that could be our initial entry point.
But selling one manually built environment only proves that a customer will pay for an environment. It doesn’t validate the most important part of our business: whether we can continuously generate high-quality environments at scale.
The deeper pain point is not creating the first set of tasks. It is what happens after the model trains on them.
“Can you continuously generate new environments and tasks as the model improves? How quickly can you do it, and is the quality comparable to environments designed by human experts?”
Read the business case and early result
Once post-training is complete, many of those tasks become too easy. The customer immediately needs a new set of tasks at the model’s new capability boundary. Today, it can take three to six months to create, implement, and validate the next environment. That slow iteration cycle is also why each environment can cost $40,000 to $50,000.
What customers really want to know is:
“Can you continuously generate new environments and tasks as the model improves? How quickly can you do it, and is the quality comparable to environments designed by human experts?”
That is what we are validating with researchers and engineers at xAI and Reflection AI.
In one of our first experiments, we generated 2,000 domain-specific environments and used them to train a 14-billion-parameter model. In our initial results, the trained model outperformed a much larger DeepSeek model on the MCP benchmark.
We are now presenting those results to AI labs as evidence that automatically generated environments can produce meaningful improvements—not just save time and money.
Early internal result; broader independent evaluation is still in progress.
Build at the capability boundary