Microsoft WebWright: An AI Web Automation Agent Architecture Based on GPT-4o and Playwright

Microsoft WebWright is a new AI web automation agent architecture developed by VenkateshDoijode (source: GitHub repo microsoft-webwright-example, released on 2026-06-07). It leverages GPT-4o and Playwright, with core design principles: separating browser sessions from agent logic to enable reusable, reflective web automation workflows. Key features include disposable browser sessions, self-reflection mechanisms, and modular skill-based task handling.

Traditional web automation agents often use a single browser session, leading to issues like being trapped in specific page states, difficulty handling complex multi-step tasks, and generating non-reusable code. WebWright addresses these gaps by decoupling the agent from browser sessions, enabling more flexible and reliable task execution.

WebWright uses a three-component architecture:

1. Runner: Orchestrates the agent loop (max 15 steps default), logs steps as JSONL, and saves successful results.
2. Model: Interacts with GPT-4o via structured JSON responses (thought, action, done fields) for transparent decision-making.
3. Environment: Handles command execution, Playwright script writing, screenshot capture, and observation formatting.

Key innovation: Disposable browser sessions—on-demand new sessions for each interaction, capture screenshots only when needed, retry failed scripts without state lock-in, and persist all products (scripts, logs, screenshots) in the workspace. Benefits: improved reliability, reusability, debugging ease, and resource efficiency.

A practical example: comparing HKG-CJU round-trip flights (2026-08-08 to 14, budget 20k HKD, economy, 3 options). Execution flow:

1. Receive task parameters → load google-flights-comparison skill → open Google Flights.
2. Capture initial fares → find cheapest direct option → pair outbound/inbound trips.
3. Identify practical options (avoid early departures) → check booking sources → compare third option.
4. Generate report and structured data.

Total duration: ~6 minutes. Outputs include logs, screenshots, flights_report.txt, and flights_data.json.

Self-reflection: Post-task, GPT-4o evaluates task success, key checkpoints, and provides improvement suggestions (saved as structured JSON).

Technical details:

• Skill system: Encapsulates task metadata, workflow steps (modular for reuse).
• Product management: Workspace stores reports, data, screenshots, logs; successful runs are copied to final_runs directory.
• Dependencies: openai (GPT-4o), playwright (browser automation), python-dotenv (env vars), requests (HTTP calls).

WebWright applies to:

• Price monitoring (e-commerce sites).
• Data scraping (structured data from dynamic pages).
• Form filling (multi-step submissions).
• Test automation (reusable end-to-end scripts).
• Workflow automation (repeatable web tasks).

Extensibility: Define new skills (Python files with task metadata/workflow) to support new tasks, leveraging modular browser operation patterns.

WebWright represents a new direction in web automation—solving traditional pain points via session-agent separation, reflection, and reusability. For developers: an extensible framework to turn natural language into executable tasks. For researchers: a model for reliable, transparent AI agents. Future: Wider applications as LLMs advance, driving smarter, more flexible automation.