initiative discovery

Autonomous Development Team

Shiv Yadav Updated 2026-03-11 apex code-factory
autonomous code-factory claude-mpm ai-engineering h1-2026

Initiative: Autonomous Development Team

What This Is

Build a software factory where tickets go in and production-ready pull requests come out. Humans review and merge — the system writes the code. Inspired by Stripe's Minions architecture (1,000+ PRs/week), built on Claude MPM and APEX.

The Core Specification

The factory is defined by a single living document: SYSTEM_PROMPT.md

This prompt configures Claude as a factory operator. It defines the 8-phase execution loop, error classification, merge conflict strategy, inter-agent data contracts, checkpointing, token budgets, and the complete operating model. Both product and engineering own it — product defines what the factory can promise, engineering defines how the factory delivers.

Why This Bet

Every hour an engineer spends writing boilerplate code, fixing lint errors, or rebasing branches is an hour not spent on architecture, design, and customer problems. The factory doesn't replace engineers — it shifts them from code writers to code reviewers and system designers.

The competitive advantage is not the model. The model is a commodity. The advantage is the harness — context engineering, deterministic gates, feedback loops, error classification, and tool integration.

Discovery Plan

  1. Stand up the factory on a single repository with S-complexity tickets
  2. Measure: PRs merged, first-pass CI rate, escalation rate, token cost per PR
  3. Expand to M-complexity tickets once S-ticket metrics stabilize
  4. Iterate the system prompt based on failure patterns (the prompt is the product)

Key Dependencies

  • Claude MPM: Agent framework, 47+ specialized agents, session management, circuit breakers
  • mcp-vector-search: 17 MCP tools for semantic code search, knowledge graphs, AI-powered review
  • APEX: Product pipeline that produces the tickets the factory consumes
  • GitHub Actions: CI pipeline the factory submits PRs against

Success Criteria

Metric Baseline Target
PRs merged / week 0 (manual) 50+
First-pass CI success N/A > 70%
PR approval without revision N/A > 60%
Escalation rate N/A < 20%
Token cost per S-ticket PR Unknown Stabilized