Agent Evaluation: Why Simulated Environments are the New Frontier for Data

The development of AI agents introduces a fundamental shift in how we must approach data for training and evaluation. While LLMs are primarily trained on static datasets to master language and reasoning, agents must be validated on their ability to act—to use tools, interact with interfaces, and execute tasks.

This leap from passive knowledge to active performance means that traditional data generation and evaluation pipelines are no longer sufficient. At Toloka, our work with teams building advanced agents has made it clear: success in the agentic era requires a move from curated datasets to high-fidelity, simulated environments.

The Limitation of Datasets for Action-Oriented Models

The standard process for LLM data generation—taxonomy design, dataset creation, and human verification—is well-established. It’s effective for teaching a model to classify, summarize, or generate text.

However, an agent's capabilities cannot be fully assessed with static input-output pairs. An agent designed for corporate workflows doesn't just need to know about Salesforce; it needs to demonstrate it can log in, pull a specific report, and transfer that data to a spreadsheet. Its performance is tied to an entire workflow within a specific context.

This necessitates an additional layer in the data pipeline: environment creation. The quality of an agent is inseparable from the quality of the environment it's tested in. We see three primary categories of agents taking shape, each requiring a distinct type of environment.

1. Generalist Agents: Interacting with Computer Environments

These agents are designed to operate a computer much like a human does, using a browser, file system, and terminal to execute complex command sequences. Evaluating them requires an environment that can replicate the intricacies of a real desktop, including its applications and potential failure states.

For a recent project with a leading LLM developer, we were tasked with red-teaming their general computer-use agent. In collaboration with cybersecurity experts, we mapped out a detailed taxonomy of potential attack vectors and then constructed a sandboxed environment containing manipulated websites, corrupted files, and other tailored challenges. This allowed us to systematically evaluate the agent's decision-making logic and safety protocols in a reproducible and controlled manner. For more details about this project, check out our recent case study.

2. Enterprise Agents: Navigating Corporate Tooling

A common application for agents is to automate workflows within a company's software stack (e.g., Google Workspace, Salesforce, Jira, Slack). The challenge here is not just tool use in isolation, but the orchestration of tasks across tools.

To address this, our engineering team developed InnovaTech, a virtual company designed as a high-fidelity testbed for enterprise agents.

InnovaTech is a pre-configured digital twin of a functioning organization, complete with virtual employees, departmental structures (Marketing, HR, Finance), and an active project history. It provides integrated access to standard enterprise tools and can be customized with client-specific software or requirements. This allows us to move beyond simple API calls and test realistic, multi-step scenarios: "Draft a project update in Google Docs based on the latest Jira tickets and share the summary in the #engineering Slack channel." This provides a safe, contained space to debug agent behavior and ensure reliability before it touches a real corporate system.

3. Specialist Agents: Mastering Industry-Specific Workflows

The third category involves agents tailored for specific industries, such as coding assistants, financial analysts, or travel booking agents. These require deep domain knowledge and fluency with specialized tools and protocols.

The growing interest in industry-specific benchmarks reflects this need.

• SWE-bench, for instance, evaluates coding agents on their ability to resolve real-world GitHub issues. We have worked with clients to develop custom, more difficult evaluations that test capabilities beyond the public benchmark. (See our case study on building a custom SWE-bench evaluation.)

• TAU-bench focuses on agent behavior in complex retail and airline scenarios, emphasizing long-term interactions and adherence to domain-specific rules.

The trend is clear: developers require testbeds that mirror the specific operational realities of their target industry. We are increasingly focused on building these custom benchmarks, as it's become a critical step for validating agent performance and safety before deployment.

The Next Engineering Challenge is the Environment

As AI systems become more autonomous, the methods we use to validate them must evolve. For agentic AI, this means treating the environment not as an afterthought, but as a core component of the development and testing lifecycle. A robust agent is one that has been pressure-tested in a world that looks like the one it will eventually operate in.

At Toloka, we are building the flexible, end-to-end infrastructure to support this shift—from generating complex, interaction-based data to environment simulation and agent evaluation.