L3:agent-orchestrated-delivery · AI Wing · Featured

Spec-driven delivery with AI coding agents

Written specs and constitutions guide parallel AI coding agents, with human review required at every merge.

Role: Workflow originator
Period: Mar 2026 to Present

Results

9+ repos
Repositories developed through the spec-first workflow
Based on repository analysis; other engineers have not yet adopted this workflow.
Verified
800+ PRs
Merged pull requests during 2026
A throughput measure only. AI agents wrote much of the code under my specs and review.
Verified

Problem

AI coding agents can produce code faster than it can be reviewed. The workflow needed clear traceability from requirements to implementation and a review gate for every merge.

What I did

I introduced GitHub Spec Kit with versioned constitutions and numbered specs linking task, acceptance and business-rule IDs.
I wrote enforceable agent instructions and dated memory logs.
I ran agents in parallel worktrees and integrated their changes through merge trains, requiring CI gates and human review for every merge.

Architecture

HumanMeWrites specs and reviews changes
DataSpec Kit specs and constitutions
AICopilot and Claude Code agentsParallel worktrees
ServiceCI gates
ExternalMain branch

Data flow

Me to Spec Kit specs and constitutions (write)
Spec Kit specs and constitutions to Copilot and Claude Code agents (tasks)
Copilot and Claude Code agents to CI gates (pull requests)
CI gates to Me (review gate)
Me to Main branch (merge)

Tech stack

GitHub Spec Kit
GitHub Copilot
Claude Code
GitHub Actions