Inspiration
AI coding agents can move from idea to implementation quickly, but speed alone does not create trustworthy engineering. Teams still need to understand scope, preserve decisions, link requirements to verification, and prevent unsafe or undocumented changes from reaching delivery.
AEGIS was created to make agentic engineering accountable rather than opaque.
What it does
AEGIS is a reusable governance layer for AI-assisted software development. It guides work through explicit stages: intake, intent alignment, requirements, checks, implementation, validation, and delivery.
It records evidence and decisions, checks traceability between requirements and tests, audits working-tree scope, detects secrets, and applies guardrails across Codex, Claude Code, Pi, OpenCode, and Qwen Code.
WAEGIS is the companion offline-first wizard. It deploys and configures AEGIS in a new or existing repository, then supports analysis, health checks, repair, and safe upgrades through a manifest-driven lifecycle.
Together, they turn an agentic workflow into something a team can inspect, reproduce, and trust.
How we built it
The project is implemented primarily in Python and designed to be offline-first and dependency-light. AEGIS ships as a copyable repository payload with a local CLI, hooks, shared skills, workflow configuration, and multi-agent adapters.
WAEGIS adds guided new, adopt, analyze, configure, doctor, repair, and upgrade journeys. Its lifecycle uses explicit plans, verification steps, journals, manifests, drift detection, rollback protection, and deterministic tests.
During OpenAI Build Week, I used Codex with GPT-5.6 to substantially extend the AEGIS and WAEGIS lifecycle, strengthen cross-platform and multi-agent guardrails, expand test coverage, and prepare the project for real repository adoption.
Challenges we ran into
The central challenge was preserving developer control without reducing the system to a static checklist. The workflow needed to be strict enough to prevent scope drift, unsafe writes, stale requirements, and secret exposure, while remaining practical for real engineering work.
Another challenge was delivering the same governance model across different agent tools without duplicating the core workflow or relying on cloud services.
Accomplishments that we're proud of
- A practical governed workflow for agentic development, not just a prompt collection.
- Traceability from user intent and requirements to checks, evidence, and delivery.
- Safe adoption of existing repositories, including component mapping and baseline review.
- Offline-first deployment, diagnostics, repair, and upgrade support.
- Guardrails and adapters across multiple coding-agent environments.
- A runnable, tested developer tool that teams can install and evaluate locally.
What we learned
Agentic development becomes more valuable when it leaves evidence behind: why a change was made, what it affects, how it was verified, and which constraints still apply.
The goal is not to make engineering autonomous. It is to make human-led engineering faster, more consistent, and easier to trust.
Log in or sign up for Devpost to join the conversation.