+ An autonomous engineering operating system that converts Jira issues into tested,
+ reviewed pull requests across multiple repositories — with a single command.
+
+
+
Core skills: built & tested
+
Orchestrator: implemented
+
Cloud deployment: remaining
+
+
+
+
+
+
The Vision
+
+
+ "Encode the entire software delivery lifecycle as composable Claude Code skills,
+ then orchestrate them autonomously across repositories — so a 4-person team
+ can ship at the pace of 40, with 80 % of actions requiring zero human input."
+
+
+
+
+
80/20
+
Autonomy model — 80 % auto, 20 % strategic checkpoints
+
+
+
~85 %
+
Average autonomy across all skills
+
+
+
40 → 3
+
Manual prompts replaced by strategic checkpoints per issue
+
+
+
8×
+
Sprint throughput vs sequential delivery (8 parallel agents)
+ A single command (/autonomous-implement ABI-123) runs the full SDLC without human
+ intervention until the plan-approval and PR-review checkpoints.
+
+
+
Jira Issue
+
›
+
Fetch & branch
+
›
+
Research codebase
+
›
+
Create plan ✋
+
›
+
Generate evals
+
›
+
Implement
+
›
+
Run evals (3× retry)
+
›
+
Create PR ✋
+
›
+
9-agent code review
+
›
+
Update Jira
+
+
✋ = strategic human checkpoint · Everything else is autonomous · Typical time: 3–8 min / issue
+
+
+
+
+
What Has Been Built
+
+
Autonomous end-to-end skills
+
+
+ AUTO
+
+
/autonomous-implement
+
Full SDLC for a single Jira issue. Accepts --context-file for orchestrator-injected knowledge.
+
+
+
+ AUTO
+
+
/autonomous-sprint
+
Parallel multi-issue sprint. Up to 8 concurrent agents in isolated git worktrees — 8× throughput.
+
+
+
+
+
Core development skills
+
+
+ GUIDED
+
+
/create-plan
+
Interactive planning from a Jira ticket. Parallel research agents, section-by-section approval, saved to specs/.
+
+
+
+ GUIDED
+
+
/implement-plan
+
Phased implementation of an approved plan. Human verification pause between phases.
Browser-driven QA. Reproduces bugs with video evidence, files 5–10 Jira issues per session.
+
+
+
+
+
+
+
+
Orchestrator Layer
+
+ A thin Python workspace layer (~500 lines) that injects repository knowledge and routes issues to the
+ correct repository before handing off to the Claude Code skills. It does not reimplement the SDLC.
+
+
+
+
Router
+
Three-strategy routing: Jira component field → keyword analysis of description → repo: label prefix. Falls back to runtime.
+
+
+
KnowledgeEngine
+
Loads per-repo knowledge (architecture, patterns, conventions, dependencies) and Foundations standards into a single context file injected as --context-file.
+
+
+
Executor
+
Creates a temp knowledge context file, invokes /autonomous-implement via subprocess or Skill tool. Handles single-repo and multi-repo flows.
+
+
+
Planner
+
Produces a TaskGraph with per-repo tasks and inter-repo dependencies. Step sequences vary by issue type (Story / Bug / Task).
+ Each of the 5 active repositories has a knowledge/repositories/<name>/
+ directory containing four Markdown files extracted from the repo's own docs:
+
+
+
+
+
architecture.md
Component map, data flows, ADR index
+
patterns.md
Coding patterns and idioms
+
conventions.md
Naming, structure, style rules
+
dependencies.md
Key libs, versions, constraints
+
+
+
+
+
+
Foundations standards
+
+ Platform-wide standards extracted from em-foundations and applied to every issue:
+
Architecture, patterns, conventions, dependencies extracted and in knowledge/. Auto-sync script watches git hash of source docs.
+
+
+
+ ✅
+
+
Workflow engine for parallel sprints
+
autonomous-sprint.js — three-phase Setup / Audit / Implement with isolated git worktrees per agent. Up to 8 concurrent.
+
+
+
+ ✅
+
+
Prompt templates for specialist agents
+
Architect, Planner, Backend, UI, Evaluator, Reviewer — parameterised Markdown templates used by the orchestrator Planner.
+
+
+
+ ✅
+
+
Repository adapters (runtime + semi)
+
Abstract RepositoryAdapter base with concrete implementations covering build, test, lint, format, branch, architecture, patterns.
+
+
+
+ ✅
+
+
Packaged as a Claude Code plugin
+
Installable via /plugin install em-software-factory@em-plugins. Plugin manifest declares all skills, hooks, and MCP config.
+
+
+
+ ✅
+
+
Security & secret controls
+
gitleaks scan script, settings block env file writes, MCP env vars loaded from shell (never committed).
+
+
+
+
+
+
+
+
What Remains: Cloud Deployment
+
+
The one open chapter — running agents and the harness in cloud
+
+ Everything described above runs locally inside Claude Code (desktop or CLI).
+ The next step is lifting the harness and agents into a persistent cloud environment so the factory
+ can run unattended, scale horizontally, and integrate with the team's CI/CD pipeline.
+
+
+
+
+
Managed Agent hosting — deploy skills as server-hosted Claude Managed Agents so they run without a developer's laptop
+
Anthropic's Managed Agents API provides a hosted sandbox with persistent context and tool access
+
+
+
+
+
Orchestrator service — wrap the Python orchestrator as a long-running service (container / Lambda / Cloud Run)
+
Needs: Jira webhook trigger → route → invoke skill → report result, all without a human in the loop
+
+
+
+
+
Secure credential delivery — Jira/GitHub tokens, MCP config, and repo SSH keys in cloud-native secret store (Vault / AWS Secrets Manager)
+
Currently loaded from developer shell env vars; needs K8s-style injection for headless runs
+
+
+
+
+
Repo access & worktree storage — agents need authenticated git clone and writable ephemeral workspace per task
+
Options: ephemeral containers per agent, shared PVC with worktree isolation, or GitHub Codespaces-style environments
+
+
+
+
+
Observability & reporting — structured logs per sprint, per agent, per issue; reporter module surfaces success/partial/failed counts
+
Reporter module exists in orchestrator but needs a sink (Datadog, Grafana, or Jira comment)
+
+
+
+
+
Air-gapped packaging — for semiconductor customers who cannot call external APIs
+
The Foundations ADRs already mandate air-gapped operation; the factory itself must be deployable without internet access