Agents

The Crew system consists of 10 specialized agents organized into phases. Each agent receives a task and shared context, performs its work, and contributes results back to the context for downstream agents.

Agent Overview

Agent	Phase	LLM Required
Impact Analyst	understand	No
Cross-Impact	understand	Yes
Regression Advisor	understand	No
Strategist	strategize	Yes
Test Designer	design	Yes
Coverage Evaluator	design	Yes
Generator	generate	Yes
Executor	execute	No
Healer	heal	Yes
Explorer	explore	Yes

Understand Phase

Impact Analyst

Maps changed files to impacted route families using the knowledge layer, dependency graphs, and traceability data. Purely deterministic — no LLM calls. Produces the foundation that all other agents build on.

Cross-Impact

Detects ripple effects across route families. When a change to “channels” also affects “messaging” through shared components or APIs, this agent identifies those cross-family links.

Regression Advisor

Scores regression risk using flaky test history, calibration data, and file-pattern heuristics. Flags areas where past instability suggests extra test attention.

Strategize Phase

Strategist

Decides the approach for each impacted flow: full-test (comprehensive coverage), smoke-test (basic verification), or skip (low risk, no action needed). Assigns priority levels and provides rationale for each decision.

Design Phase

Test Designer

Produces structured TestCase[] objects with type classification, preconditions, step-by-step instructions, expected outcomes, and rationale. Covers 9 test categories: happy-path, edge-case, boundary, negative, state-transition, race-condition, permission, accessibility, and performance.

Coverage Evaluator

Reviews the test designs against the impact results and identifies remaining coverage gaps. Ensures the design phase produces adequate coverage before moving to generation.

Generate Phase

Generator

Converts test designs into Playwright spec code. Uses the discovered API surface (page objects, custom helpers) to generate specs that use real project methods rather than hallucinated ones.

Execute Phase

Executor

Runs the generated specs and collects pass/fail results, error messages, and screenshots. Provides execution data to the healer for failure analysis.

Heal Phase

Healer

Analyzes test failures from the executor, diagnoses root causes (selector changes, timing issues, API mismatches), and patches the broken specs. Can iterate multiple times to resolve cascading issues.

Explore Phase

Explorer

Autonomous browser agent that navigates the application, tries edge cases, and records findings. Used by the impact-gate-qa for browser-based exploratory testing.