Quick answer
Prefer a stable, authorized API when it covers the business operation. Choose browser automation for a web-only workflow when the browser boundary is sufficient and supported. Choose WorkSpaces Applications agent access when a necessary Windows or legacy application has no suitable interface and needs a managed desktop session. Keep a human workflow when consequence, variability, or evidence makes automation unsafe or uneconomic.
The strongest architecture may be hybrid: API for status and writes, MCP forwarded tools for structured operations, desktop control for the one legacy gap, and human approval before consequence.
Decision matrix
| Criterion | Direct API | Browser automation | WorkSpaces agent desktop | Human operation |
|---|---|---|---|---|
| Best fit | Supported structured operation | Web-only interface | Windows or mixed desktop apps | High ambiguity or consequence |
| Contract | Versioned schema and status | DOM, accessibility tree, browser behavior | Screenshot, input tools, session and app UI | Procedure and training |
| Identity | Token or signed request | Browser session plus backend identity | IAM MCP request, streaming/SAML session, app login | Named human account |
| Verification | Response plus system read | DOM and system record | Screen, logs, screenshots, system record | Peer review and audit |
| Main fragility | API change, quota, partial failure | Page and auth changes | Image, screen, app, session, and model changes | Throughput and consistency |
| Cost drivers | Calls, compute, engineering | Browser runtime, retries, model | Fleet, agent session, user, model, storage | Labor and queue time |
Do not choose by demo speed. Compare accepted outcome, correction effort, operational ownership, and exit cost.
Prefer the API when it is suitable
An API gives typed inputs, explicit errors, stable identifiers, idempotency options, and direct status reads. It is normally the best route for financial, identity, inventory, messaging, and other consequential writes—provided it is supported, authorized, sufficiently complete, and has the needed regional and data controls.
An API does not solve bad authorization or unsafe business rules. Bind access to the verified user and resource, validate inputs, set budgets, and reconcile ambiguous timeouts.
Choose browser automation for web-only scope
Browser automation avoids provisioning a full Windows image when all required applications are web-based. DOM or accessibility semantics can be more reliable than image coordinates, but page changes, authentication, extensions, downloads, cross-origin content, and bot controls still create risk.
Use a dedicated browser profile, restrict destinations, protect cookies and tokens, and test the supported browser surface. Do not circumvent application terms or security controls.
Choose WorkSpaces when the desktop is the requirement
WorkSpaces agent access is differentiated by the managed streaming desktop and MCP service. It can reach Windows applications without modifying them and supports computer input, vision, optional screenshots, forwarded MCP tools, IAM signing, CloudTrail, CloudWatch, and observer modes.
The tradeoff is a broader operating surface: fleet and stack, Windows image, applications, network, WorkSpaces Agent, streaming session, IAM, model, screenshots, and UI behavior. Current service limitations also exclude several fleet types and VPC endpoints.
Read the GA overview and regions and pricing guide before selecting it.
Use a hybrid workflow deliberately
Break the process into steps and choose the strongest interface for each:
- retrieve request and entitlement through an API;
- open the legacy application in a dedicated WorkSpaces session;
- use computer vision only to locate the required record;
- use a forwarded MCP tool or API for a deterministic transformation where available;
- present the exact proposed change to a human;
- commit through the API if supported, otherwise through the desktop with an operation ID;
- read the system of record and reconcile;
- close and expire the streaming session.
This reduces screen dependence without forcing premature application modernization.
Run a matched pilot
Give each serious option the same 20–50 representative tasks, identities, failure cases, and acceptance rubric. Measure successful accepted completion, false success, duplicate attempt, human intervention, latency, cost, security exceptions, accessibility, audit completeness, maintenance hours, rollback time, and user impact.
Stop if the system cannot distinguish failure from unknown outcome, cannot bind authority to a resource, or cannot recover from a UI or session change.
Use the governance guide for operating controls and the AWS agent desktop readiness checklist for the final go/no-go.
Frequently asked questions
API or desktop?
Prefer an API for supported deterministic operations. Use desktop access for a necessary legacy UI gap that can be bounded and verified.
Browser or WorkSpaces?
Browser automation may be simpler for web-only work; WorkSpaces fits Windows and mixed desktop applications that need a managed streamed environment.
Can the methods be combined?
Yes. A hybrid often minimizes UI risk while retaining access to legacy steps.
Official sources
- AWS: WorkSpaces agent access overview
- AWS: agent access capabilities and limits
- AWS: managed MCP server
Source check: August 12, 2026. Verify the actual application’s supported APIs, browser requirements, terms, identity model, and WorkSpaces eligibility before choosing an architecture.