A training gym for AI agents.

The idea came from six months spent building enterprise post-training environments.

One project involved helping a financial-services team train an AI model on complex 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 domain experts and identify around 30 tasks they perform regularly. Then we manually created Docker-based sandbox environments containing mock spreadsheets, mock business-system 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 across payments and legal workflows.

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 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 validate whether the request makes sense and resembles work a real employee might ask an AI agent to do. Any findings stay attached to the task for review; they do not block or hide the executable release.

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.

Once post-training is complete, many of those tasks become too easy. The next round needs a new set of tasks at the model’s updated capability boundary. A manual process can take months to create, implement, and validate the next environment.

Continuous generation shortens that iteration cycle and lets each new task set track the model’s current capabilities.

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.

Early internal result; broader independent evaluation is still in progress.