Plato raises $16 million to build simulated workplaces for AI agents

Bridging the gap to AI agents

Plato has secured $16 million in Series A funding to address a specific hurdle in the evolution of artificial intelligence. While current chatbots are proficient at generating convincing answers, they often struggle with the actual execution of complex jobs. The company aims to provide a space where models can learn to navigate business decisions and use various software tools effectively.

The funding, led by Nava Ventures, brings Plato’s total capital to $19 million. This investment comes at a time when the industry is moving toward autonomous agents capable of handling tasks like purchasing, support cases, and accounting workflows. To do this successfully, an agent must be able to reconcile different pieces of information while respecting specific constraints.

The core philosophy of the company is that mastering complex work requires experience. As co-founder and CEO Robert Farlow noted, "Mastering real complex work comes from experience doing those tasks." By creating a simulated environment, Plato allows models to gain this experience in a controlled setting before they are ever deployed into live, high-stakes systems.

How simulated workplaces work

Plato functions as an applied research lab that converts real-world software and data into training environments. Developers can create sessions that include multiple simulators, such as replicas of GitHub, Jira, and Stripe. This allows an AI agent to complete a single task that might require interacting with several different platforms simultaneously.

A key feature of these simulations is the ability to reset the workspace. In a live business system, an agent making a mistake might spend actual money or permanently alter a customer record. In a Plato simulation, the environment can be wiped and reused, allowing developers to run the same problem multiple times to see how the agent handles different variables.

The company works with domain experts to recreate specific jobs in fields like healthcare administration, finance, and people operations. They use software replicas and mock data to build these environments. For example, one task requires an agent to prepare a profit-and-loss statement using simulated spreadsheets, emails, and banking tools, with separate verifiers checking for correct currency conversions and accounting treatments.

Training through reinforcement learning

AI labs can use these simulated environments for a process called reinforcement learning. In this setup, models attempt various tasks and learn from the feedback they receive across repeated experiences. This method is essential for teaching models how to handle unexpected situations that often make professional work difficult in the real world.

The quality of the feedback is critical to the success of the training. If a simulation only rewards a finished task without checking if the agent followed the rules, the model might learn to prioritize speed over accuracy. Plato builds evaluations that check both the final outcome and the specific path the model took to get there, ensuring it respected all instructions.

The company also monitors for "reward hacking," which is when a model finds a shortcut or exploit in the simulation to get a high score without actually doing the work correctly. By focusing on high-fidelity replicas, Plato ensures that the strategies an agent learns in the simulation will actually translate to the decisions it faces in a real-world office environment.

The road to real-world deployment

With the new capital, Plato plans to expand its library of work environments and the experience available to developers. They intend to pay professionals to recreate more of their daily jobs into these simulated environments to increase the variety of tasks available for training. This helps build a more diverse and robust set of training data for the next generation of AI agents.

The company is also looking toward the future by planning to deploy agents on real tasks to collect additional experience. They are developing richer simulations that involve longer workflows and interactions with simulated people. This progression aims to create a more nuanced understanding of how human-AI collaboration works in a professional setting.

Plato's work sits within a broader trend of investment in agent-training infrastructure. By providing a way to verify workflows and practice complex tasks, the company is betting that the value of AI will depend on how well these models can be prepared for the rules of professional work. Their goal is to provide a reliable bridge between a model's ability to reason and its ability to act.