What Makes an Agent
An agent is more than a prompt. It loops and takes actions toward a goal.
Agent = model + tools + goal + memory + loop
AI Agents Cheat Sheet
A compact reference for the agent loop and the practical agent patterns we teach for QA and automation work.
The Agent Loop
This is the simplest mental model to remember.
Observe current state Think about next best action Act using a tool Observe result Repeat until goal is met
Common Agent Roles
These show up repeatedly in testing-oriented systems.
Planner: understands the app and creates a plan Generator: writes or updates test code Healer: diagnoses and fixes broken tests
Useful Agent Tools
Agents become valuable when they can interact with real systems.
- browser automation - file read/write - terminal commands - issue trackers - HTTP APIs - vector search / RAG
Memory
Memory helps the agent retain context across steps or sessions.
Short-term memory: - current conversation / task state Long-term memory: - saved facts - previous outcomes - user preferences
Good Agent Prompts
Agent prompts should define role, objective, limits, and when to stop.
You are a QA planning agent. Explore the target application. Identify high-risk flows. Do not make destructive changes. Return a prioritized test plan in markdown.
Guardrails
Agents need boundaries, especially when tools can write or execute.
- limit tool scope - require confirmation for risky actions - cap retries - restrict secrets access - log actions and outputs
Evaluation
Agent quality should be measured by outcome quality, not how smart the logs sound.
Measure: - task completion - correctness - efficiency - unnecessary actions - safe behavior - repeatability
QA Use Cases
These are realistic places where agents can assist testing teams right now.
- test planning from requirements - code review for flaky tests - CI failure triage - test data generation - bug summary writing - healing locator changes