Plato

Plato was founded in 2025 by CEO Robert Farlow and CTO Pranav Putta.⁠[1]

A Gradient Ventures research article identified Plato as a computer-use environment supplier and discussed its GFRL experiment in prompt and context optimization. This approach uses environment feedback to improve inputs rather than directly updating model weights, illustrating how simulations can support different optimization methods.⁠[2]

Plato introduced Web Clone Bench, testing whether models can reconstruct nine reference applications by inspecting running products. It covers interfaces such as Gmail, Figma and Slack and evaluates both appearance and behavior. Models run in their vendors’ coding harnesses, so results describe this specific task and setup.⁠[3]

Computer Work Val v1.0 contains 300 tasks across 10 domains and 36 roles, using 210 simulated apps. Practitioners recreate work with mock identities and accounts. Verifiers check resulting states, while reviewers examine realism, fairness and weaknesses in grading.⁠[4]

Plato raised a $16 million Series A led by Nava Ventures, with Gradient Ventures, Transpose Platform and Mustafa Suleyman participating, bringing total funding to $19 million. Nava founder Freddie Martignetti joined the board; proceeds would expand its simulated-work library.⁠[1]

The platform distinguishes prebuilt application simulators from running environments, with sessions grouping multiple environments. Testcases define initial artifacts, prompts and scoring settings. Evaluation can check database or file changes, structured outputs, or both, making task completion programmatically verifiable.⁠[5]

Environments can use simulator examples, pinned artifacts or custom-resource VMs. Documentation notes that example versions can change, making fixed artifact IDs preferable for reproducible runs. Each environment has an alias and internal hostname and supports individual state inspection, reset and shutdown.⁠[6]

Sessions coordinate environment lifecycles. VMs terminate after a default 30-minute timeout, which can be configured, and documentation instructs users to close sessions when finished. Only testcase-created sessions carry scoring configuration; arbitrary environment combinations do not automatically provide evaluation criteria.⁠[7]

The Core SDK provides Python access; the current quickstart requires Python 3.10 or later and plato-sdk-v2 2.61.4 or later. Personal and organization API keys can connect to the main service or a dedicated tenant. Legacy keys are deprecated, and documentation warns against committing environment-access keys.⁠[8]

Plato’s computer-use environment provides a Linux desktop, Chrome and a filesystem. Agents interact through screenshots, mouse and keyboard actions, shell commands and file operations, with live viewing for observation. Desktop and application simulations share session interfaces and can be combined for cross-tool tasks.⁠[9]

Plato describes itself as an applied research lab that builds simulations around partners’ software, data and target capabilities. Its project-inquiry process addresses teams developing and evaluating agents, with a focus on training and validation environments.⁠[10]

The privacy policy describes collecting inquiry details for service delivery, communication and improvement. The service terms assign platform content and functionality to Plato Technologies, Inc. while preserving users’ ownership of submitted data. These public policies do not themselves establish a particular security certification.⁠[11][12]