Forge

// 01 - PROBLEM

Most "AI agent" demos are a prompt and a spinner. Real software delivery needs a team: something that plans a feature, simulates the implementation, writes a test plan, runs quality gates, and asks a human before anything ships. Forge models that whole control loop.

// 02 - APPROACH

• ▸A typed runtime engine drives a 9-step pipeline from feature request to PR-ready artifact.
• ▸Four specialist agents own their slice: planner, implementer, test author, reviewer.
• ▸Every gate produces a score, not a vibe - releases are blocked below threshold.
• ▸An explicit human-in-the-loop approval gate sits before the final artifact.

// 03 - ARCHITECTURE

Mock mode as a first-class seam

The LLM and database sit behind clean adapters, so the entire platform runs with zero API keys. Reviewers can run it in one command; going live touches config, not app code.

Zod schemas as the single source of truth

Every agent boundary is a validated contract. Malformed agent output fails loudly at the seam instead of corrupting downstream steps.

// 04 - PRODUCTION-GRADE

• ✓Per-step observability: tokens, cost, retries, and logs for every pipeline step
• ✓Scored quality gates that block, not warn
• ✓Human approval gate before artifact creation
• ✓Runs with zero API keys in mock mode; CI on every push

// 05 - ARTIFACTS

• github.com/mikulgohil/se-agent-platform →