供资
Plato 融资 1 600万系列A，用于建设AI Agents的模拟工作场所
在 Google 上的首选来源中添加 Unite.AI
柏拉图拥有融资 1600万美元A系列资金，用于建设模拟工作场所，AI模型可以使用软件进行练习，导航商业决策，并在运行于直播系统之前完成复杂的工作。 Nava Ventures领导了该轮比赛，参与的有Transpose平台的Gradient Ventures和Microsoft AI的首席执行官Mustafa Suleyman.
在Gradient Ventures领导的种子回合之后，带来 Plato的融资总额达到1900万美元。 Nava Ventures的创始人Freddy Martignetti即将加入董事会。 柏拉图成立于2025年，总部位于旧金山，将利用首都扩充其工作环境和AI开发者可用经验的图书馆。
投资的目标是实际限制从聊天机到agents的移动：产生令人信服的答案和进行工作是不同的能力。 agent 处理购买，支持案件，或会计工作流程必须使用正确的工具，调节信息，尊重约束，并将基础记录留在正确的状态。
一个无需更改现场记录即可练习的地方
在真正的商业系统内训练agent可以在agent准备好之前产生后果。 采购任务可能花钱，而行政工作流程可能改变客户或雇员的实际记录。 柏拉图在包含的环境下重新创建那些工作流程，开发者可以重置和再利用。
称 Robert Farlow，柏拉图的联合创始人兼总裁，该公司的模拟旨在捕捉使专业工作困难的工具，决定和意想不到的情况。
柏拉图与域专家合作，在客户支持、人员操作、财务、会计和保健管理方面重新创造就业机会。 AI实验室可以利用由此产生的环境进行强化学习，其中模型尝试任务，从反复经验的反馈中学习，以及评价其性能。
从软件模拟器到可验证的工作流程
公司网站称柏拉图是一个应用研究实验室，将现实世界的软件和数据转化为培训和评价agents的环境。 其说明性开发者的工作流程创建了包含多个模拟器的会话，加载一个任务，运行agent，并评价由此展开。 例子结合了GitHub,Jira，和Stripe模拟器，显示了一个任务如何跨越几个工具。
这种结构给开发者一个可重复的经验单元：一个初始的工作空间，一个完成的任务，agent的动作，以及检查结果。 重置工作空间使得有可能在不积累活账户变化的情况下再次尝试同样的问题。
Plato 的 Computer Work Val v1.0 的调查报告，日期为10月27日，称贡献者公司将自己的工作片段转化为由软件复制品和模拟数据构建的环境。
其中一个例子是要求agent使用模拟电子表格、电子邮件、付款和银行工具编制季度损益报表。 单独核查者检查会计处理，正确的货币折算，交付最终档案等要求。 这种做法可能暴露在原本已经完成的工作流程中发生的一个间接错误：报告可能到达正确的接受者手中，同时仍然高估收入。
报告还介绍了有关职业专家的审查，并检查环境缺陷、提示和分级不匹配以及奖励黑客。 其重点是包含成功所需信息的任务，同时仍然需要从多个来源进行推理。
梯级系统为何重要
模拟的反馈决定了模型的学习。 如果评价奖励完成的购买，而不检查预算，则可能教agent优先完成，而不是使任务有意义的指示。
Plato 称它评价结果和模型是如何达到它的，包括它是否遵循指令和尊重的制约。 它还审查模型是否利用环境中的弱点来获得不反映正确工作的分数。
称 Pranav Putta、柏拉图的联合创始人和CTO。
公司独立的Web Clone法官组审查模型从参考环境重现网络应用的能力。 柏拉图将功能检查和人的质量保证描述为建设高可靠性复制品的一部分。 对于培训基础设施，忠诚是技术要求：缺失的行为或不切实际的捷径可以改变agent学习的战略。
更多真实世界经验
法罗之前在尤托里建造了一个注释平台，支持计算机使用模型培训。 Putta的背景包括计算机视觉，机器人，强化学习，以及MultiOn的计算机使用模型工作。
新的资金将支付专业人员的费用，以便在模拟环境中重新创造更多的工作。 柏拉图还计划将AI agents部署在真实的任务上，以收集更多的经验，并正在开发更丰富的模拟，使用更长的工作流程以及与模拟人员的互动。
该回合是在对agent培训基础设施进行更广泛的投资之后进行的，包括Labelbox对强化学习环境的资助和AI agents. 柏拉图的赌注是，模型需要扩大现实做法的供应，同时配以保护工作规则的评价。 这种基础设施的价值取决于模拟中的经验如何使agent为它将面临的外部决定做好准备。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Funding
Plato Raises $16 Million Series A to Build Simulated Workplaces for AI Agents
Add Unite.AI to your preferred sources on Google
Plato has raised $16 million in Series A funding to build simulated workplaces where AI models can practice using software, navigating business decisions, and completing complex jobs before they operate in live systems. Nava Ventures led the round, with participation from Gradient Ventures, Transpose Platform, and Mustafa Suleyman, CEO of Microsoft AI.
The financing brings Plato’s total funding to $19 million, following a seed round led by Gradient Ventures. Freddie Martignetti, Nava Ventures’ founder, is joining the board. Founded in 2025 and based in San Francisco, Plato will use the capital to expand its library of work environments and the experience available to AI developers.
The investment targets a practical limitation in the move from chatbots to agents: producing a convincing answer and carrying out a job are different capabilities. An agent handling a purchase, a support case, or an accounting workflow must use the right tools, reconcile information, respect constraints, and leave the underlying records in the right state.
A Place to Practice Without Changing Live Records
Training an agent inside a real business system can create consequences before the agent is ready. A purchasing task might spend money, while an administrative workflow could change an actual customer or employee record. Plato recreates those workflows in contained environments that developers can reset and reuse.
“Mastering real complex work comes from experience doing those tasks,” said Robert Farlow, Plato’s co-founder and CEO, in the announcement. The company’s simulations are intended to capture the tools, decisions, and unexpected situations that make professional work difficult.
Plato works with domain experts to recreate jobs in customer support, people operations, finance, accounting, and healthcare administration. AI labs can use the resulting environments for reinforcement learning, in which models attempt tasks and learn from feedback across repeated experience, as well as for evaluating their performance.
From Software Simulators to Verifiable Workflows
The company’s website describes Plato as an applied research lab turning real-world software and data into environments for training and evaluating agents. Its illustrative developer workflow creates a session with multiple simulators, loads a task, runs an agent, and evaluates the resulting rollout. The example combines GitHub, Jira, and Stripe simulators, showing how one task can span several tools.
That structure gives developers a repeatable unit of experience: an initial workspace, a task to complete, the agent’s actions, and checks on the result. Resetting the workspace makes it possible to attempt the same problem again without accumulating changes in a live account.
Plato’s Computer Work Val v1.0 research report, dated October 27, describes a validation set of 300 tasks spanning 10 domains and 36 roles, running across 210 simulated applications. The company says contributors turn pieces of their own jobs into environments built from software replicas and mock data.
One example asks an agent to prepare a quarterly profit-and-loss statement using simulated spreadsheet, email, payments, and banking tools. Separate verifiers check requirements such as accounting treatment, correct currency conversion, and delivery of the final file. That approach can expose a consequential mistake inside an otherwise completed workflow: a report may reach the right recipient while still overstating revenue.
The report also describes review by experts from the relevant occupation and checks for environment defects, mismatches between prompts and grading, and reward hacking. Its emphasis is on tasks that contain the information needed to succeed while still requiring reasoning across multiple sources.
Why the Grading System Matters
A simulation’s feedback shapes what a model learns. If an evaluation rewards a completed purchase without checking the budget, it may teach the agent to prioritize completion over the instructions that made the task meaningful.
Plato says it evaluates the outcome and how the model reached it, including whether it followed instructions and respected constraints. It also examines whether models exploit weaknesses in an environment to obtain a score that does not reflect correct work.
“We build evaluations that reflect what the job actually requires, so developers can distinguish a high score from work done correctly,” said Pranav Putta, Plato’s co-founder and CTO.
The company’s separate Web Clone Bench examines models’ ability to recreate web applications from reference environments. Plato describes functionality checks and human quality assurance as part of building high-fidelity replicas. For training infrastructure, fidelity is a technical requirement: missing behaviors or unrealistic shortcuts can change which strategies an agent learns.
Funding More Real-World Experience
Farlow previously built an annotation platform at Yutori supporting computer-use model training. Putta’s background includes computer vision, robotics, reinforcement learning, and work on computer-use models at MultiOn.
The new funding will pay professionals to recreate more of their jobs in simulated environments. Plato also plans to deploy AI agents on real tasks to collect additional experience and is developing richer simulations with longer workflows and interactions with simulated people.
The round follows wider investment in agent-training infrastructure, including Labelbox’s financing for reinforcement-learning environments and AI agents. Plato’s bet is that models need an expanding supply of realistic practice, paired with evaluations that preserve the rules of the work. The value of that infrastructure depends on how well experience inside a simulation prepares an agent for the decisions it will face outside it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.