e2e-llm-agents

Principles for stack-agnostic E2E testing with LLM agents. Specs and illustrative examples — not a working install template.

Companion to:

• the article [title TBD] — empirical case study (8 runs / 11 production bugs / 14%→95% KB-saturation curve)
• the canonical methodology page at https://webmaster-ramos.com/docs/e2e-llm-agents/

What this is

Prose specifications and light illustrative examples for an E2E methodology forged in production on a real e-commerce platform over 8 calibration runs (11 production bugs found, KB-saturation 14%→95%), then validated independently on a second stack (FastAPI + Alpine, 48.6%→91.4% in two runs).

The methodology applies the six principles from Six Principles for Agent Systems That Don't Hallucinate to the specific task of E2E test generation and healing.

What this is NOT

• Not a working install template. There is no bootstrap script. There is no playwright.config.ts. You cannot git clone this and get a runnable system.
• Not sanitised production code. Specs are written from scratch as principle descriptions; examples are written from scratch against a public demo target. The production e2e-system stays private.
• Not a Playwright tutorial. It assumes Playwright fluency. It covers the agent layer above test code, not the test code itself.

Layout

• specs/ — prose specifications: contract, four-layer structure, four agent roles, skill design, knowledge format, no-retries doctrine, heal-findings structure, lint-as-KB, metric design.
• examples/ — light illustrative artifacts written against todomvc.playwright.dev as a neutral target. Read-only reference; not runnable.

Reading order

1. The article — case-study evidence and motivation.
2. The /docs/e2e-llm-agents/ page — canonical methodology narrative.
3. specs/ — granular specification for each piece.
4. examples/ — concrete artifact shapes.

License

MIT — see LICENSE.