From Assistants to Agents Embedded in Enterprise Workflows
AI usage is rapidly shifting from simple assistants that answer questions to embedded AI agents that execute complex workflows across enterprise systems. Instead of just summarizing documents or suggesting routes, agents are now expected to make decisions, orchestrate tools, and take actions. This evolution depends on two pillars: access to accurate, domain-specific data, and structured workflows that guide multi-step reasoning and execution. Mapping platforms such as TomTom’s Agent Toolkit illustrate this direction: agents can query rich geolocation data to support underwriters, city planners, or fleet dispatchers using natural language. In this agentic era, the conversation becomes the interface, while specialized toolkits and data sources sit behind the scenes. Yet a major gap has persisted: most enterprise workflows still run through legacy desktop systems lacking modern APIs, limiting where AI agents can actually act. That is the friction AWS WorkSpaces now targets.
How AWS WorkSpaces Lets AI Agents Drive Legacy Desktop Applications
AWS WorkSpaces now allows AI agents to operate legacy desktop applications directly, using the same virtual desktops human employees use. Instead of demanding custom APIs or modernization projects, WorkSpaces gives an agent a managed desktop session reachable via a pre-signed URL. The agent authenticates through IAM, then controls the environment through computer vision integration and input simulation: taking screenshots, interpreting the interface, clicking buttons, typing into fields, and scrolling through records. To the legacy application, nothing has changed—it simply receives keyboard and mouse events. This approach dramatically broadens desktop application automation, because any software that a human can use through a graphical interface becomes accessible to AI agents. For organizations wrestling with AI agents and legacy systems, WorkSpaces reframes the question from “Can we expose an API?” to “Can we let an agent log into a desktop like a user?”
Why This Matters for Legacy Systems and Regulated Environments
Enterprises have long struggled to extend enterprise workflow automation to critical systems that predate modern integration patterns. A Gartner report cited by AWS notes that most organizations still run legacy applications with no suitable APIs, and many large firms depend on mainframe-hosted processes with limited programmatic access. Rewriting or wrapping these systems can be risky and slow, especially in regulated industries. WorkSpaces offers another path: AI agents operate within the same secure, governed desktops as human staff. Identity and access are managed through IAM; activity flows into CloudTrail and CloudWatch for audit and observability; and each agent can be given its own identity, separating automated from human actions. Because desktop application automation occurs inside isolated WorkSpaces instances, enterprises can adopt AI agents for sensitive workflows without exposing internal networks or rebuilding mission-critical legacy systems.
Computer Vision Integration Expands Automation—But at a Cost
The breakthrough here is computer vision integration: agents inspect screenshots, understand on-screen elements, and decide where to click or type. This unlocks AI agents for legacy systems that were never designed for automation, but it also introduces efficiency trade-offs. Research highlighted in the AWS announcement shows that a vision-based agent may consume vastly more model tokens than an API-based agent for the same task, and take dramatically longer to complete it. Better vision models can reduce errors per screenshot, yet they do not reduce how many screens an agent must interpret to reach the desired data. For high-value, low-volume workflows—like a pharmacy refill inside a legacy desktop system—this overhead may be acceptable, because the alternative is no automation at all. For high-volume scenarios, enterprises will likely blend API integrations where available with vision-driven desktop application automation where APIs do not exist.
Toward an Agentic Architecture that Includes Legacy Apps
The combination of WorkSpaces and MCP endpoints points toward an architecture where AI agents can orchestrate tools across both modern and legacy environments. Because WorkSpaces exposes a managed MCP server, framework-agnostic AI agents built with systems like LangChain, CrewAI, or Strands can treat desktop control as just another tool in their toolbox, alongside APIs, databases, or specialized SDKs such as geospatial toolkits. In practice, an agent might query location intelligence, reason about optimal actions, and then execute those actions by driving a legacy desktop application, all through a single conversational interface. For enterprises, this reduces the barrier to adopting agentic AI: instead of waiting for every system to be modernized, they can let agents work with what exists today. Over time, modernization can proceed selectively, while agents continue to bridge old and new worlds in enterprise workflow automation.