Slack Uses AI Agents for End-to-End Testing in Dynamic UIs

Slack Uses AI Agents for End-to-End Testing in Dynamic UIs

Slack's engineering team has deployed what it calls "agentic testing," an approach that uses AI agents to execute end-to-end tests based on high-level objectives rather than rigid, predetermined sequences. The method addresses a persistent problem in continuous delivery: tests that break when user interfaces or service structures change, even when underlying functionality remains intact.

Traditional end-to-end tests depend on stable selectors, fixed action sequences, and predictable application flows. When developers modify UI elements or relocate components, these tests fail—not because features are broken, but because the test script can no longer navigate the changed interface. The result is mounting maintenance overhead as teams spend time updating test code rather than shipping features.

How agentic testing works

In Slack's implementation, engineers define a test as a goal rather than a script. An AI agent receives the intent—for example, "send a message in a specific channel"—and determines how to accomplish it by interacting with the application's UI or API surfaces. At each step, the agent evaluates the current application state and selects the next action dynamically.

When the agent encounters minor changes, such as a relocated button or modified element structure, it attempts alternate paths instead of immediately failing. Once the objective is complete, the system validates the outcome against engineer-defined assertions. Execution traces capture the sequence of decisions and interactions, enabling teams to replay and inspect failures.

The workflow begins with passing a test intent to the agent layer. The agent performs planning, executes actions against the system under test, observes results, and iterates until reaching the objective or hitting a stopping condition. Engineers configure constraints to govern agent behavior, including limits on allowed actions, exploration boundaries, and termination conditions.

Why it matters

This approach doesn't replace deterministic testing—it complements it. Slack continues to rely on traditional tests for validating critical logic and contract correctness, particularly at the unit and integration layers. Agentic testing targets the end-to-end layer where workflows are most vulnerable to superficial changes that don't reflect actual regressions.

The cost profile currently makes agent-driven testing better suited for targeted debugging and exploratory testing rather than frequent execution in CI pipelines. Slack positions it as a tool for exploring complex UI behavior, debugging flaky workflows, and reproducing production issues—scenarios where adaptability matters more than speed.

Deterministic tests provide fast, repeatable regression validation. Agent-based execution operates from a goal-oriented model, observing state and dynamically determining how to reach the desired outcome. The two approaches serve different needs within the testing pyramid.

Implementation considerations

Observability is central to making agentic testing practical. Execution logs are structured to provide visibility into each decision the agent makes. This transparency allows teams to understand why an agent chose a particular path and to identify when adaptive behavior masked a genuine problem.

The system includes guardrails to prevent unbounded exploration or unintended actions. Engineers define what the agent can and cannot do, set boundaries for how far it can deviate from expected paths, and establish clear stopping conditions.

Details of Slack's agentic testing approach were first reported in a blog post by the company's engineering team.