An open source framework born at Disruptica.
ADRIAN is a software development framework for small, self-managing teams that produce most of their code with AI agents (Claude Code, Cursor, Aider, etc.) and use an in-app requirements bot (BugLady) as the requirements spine.
Born at: Disruptica — where we run ADRIAN daily on our own products Author: Jay (Jairo) Anaya Companion tool: BugLady — the in-app PM agent that interviews users and produces ADRIAN's Request Contracts. Disruptica's internal deployment of BugLady is codenamed Ramjet — same tool, dogfooded on our own codebases before anything ships to buglady.ai.
Software development used to be chess: study the board, plan ten moves ahead, commit to a strategy, execute it over months. Roadmaps, quarters, epics.
AI-first development is squash: move fast, react to where the ball actually is, reposition constantly. The ball is user need — it moves every day. Winning isn't predicting; it's reacting with quality at speed without loosing the north.
ADRIAN is built for squash:
- The unit of work is a Request — something a user or the business asked for, small enough to deliver in hours to days
- Priority drives strategy, not the other way around — the queue reshuffles as often as reality changes
- ADRIAN rewards change. Killing an in-flight Request because priorities shifted is not failure — it's the system working. Requests are small enough that killing one is cheap; kill and respawn beats mutating mid-flight
- Quality lives in the contract and the pipeline, not in slow ceremony — gates are automated and upstream, so speed doesn't cost correctness
ADRIAN is a framework we built for ourselves first and open-sourced second. It is not (yet) a widely-adopted industry standard. There is no widely-adopted standard for AI-first agentic teams — the field is roughly 18 months old as of mid-2026.
What ADRIAN is:
- A synthesis of converging 2025–2026 practices: request-driven contracts, agent context engineering, AI-attributed quality metrics, mutation testing, architectural fitness functions
- Battle-tested at Disruptica on a team structure of 1 PM + 1 or more Devs + 1 QA/Success, with BugLady (internally: Ramjet) as the requirements engine
- Designed to flex across unregulated and regulated domains via a compliance mode switch
- Version 0.1 — revised quarterly based on the metrics it produces
What ADRIAN is not:
- A published industry framework like Scrum or XP
- A replacement for engineering judgment
- A guarantee of quality — it is a structure that enables quality if practiced
References this draws from are listed in REFERENCES.md.
The requirements are doing less work than they need to, and review is doing more work than it can sustain.
When AI produces 85%+ of code, traditional review cannot keep up. The quality gates move upstream into the Request Contract and downstream into automated verification. Humans spend attention only on intent, judgment, and accountability.
- A user (or the PM, or the business) talks to BugLady inside the application — a chat interview, the way a product manager would ask questions
- BugLady analyzes the conversation, the codebase, business goals, and compliance posture
- The interview ends in a Request Contract — structured sections plus the interview itself, kept in BugLady as context for both the dev team and the coding agents
- The PM pressure-tests, sets priority (User Impact × Business Context), and approves
- The Request enters the queue, gets implemented by agents under Dev steering, ships through automated gates, and is verified in staging and production
The interview never goes to git. It stays in the requirements tool. Git gets the PR, which links back to the Request Contract.
| Phase | Who leads | Output | Time share |
|---|---|---|---|
| Align | PM + BugLady | An approved Request Contract | 15–30% |
| Implement | Dev + AI agents | Working code, tests, docs | 40–50% |
| Deliver | CI/CD + QA | Verified production deploy | 15–25% |
| Adapt | Whole team + agents | Lessons learned → team practice AND agent context (CLAUDE.md, BugLady prompts) |
5–10% |
The Adapt phase feeds two learners: the team (retros, quality review) and the agents (updated rules files, updated BugLady interview prompts). Both compound.
Read in this order if you're new to ADRIAN:
METHODOLOGY.md— The full framework, all four phases in detailROLES.md— What PM, Dev, and QA actually do per phaseREQUEST-CONTRACT.md— The Request Contract template (this is what BugLady produces)PHASE-GATES.md— Checklists for moving a Request between phasesFLEX-MODES.md— Compliance switches (none / SOC2 / HIPAA / PCI / gov)CLAUDE-MD-TEMPLATES.md— Agent instruction files for TS/Node, Python, GoQUALITY-DASHBOARD.md— Five metrics, thresholds, what each triggersADOPTION-PLAN.md— Week-by-week rolloutREFERENCES.md— What this draws from and what's original
Additional repo files:
CONTRIBUTING.md— How to propose changesLICENSE— MITCITATION.cff— If you cite ADRIAN in writing
BugLady is an in-app product manager that lives inside each application you produce or maintain. It captures requirements through a chat interview by analyzing the codebase, business goals, and compliance posture, then generates the Request Contract that drives ADRIAN's Align phase — including the interview itself as context for the dev team and coding agents.
At Disruptica, our internal deployment of BugLady is codenamed Ramjet. Same tool — we run it against our own products first. Every ADRIAN practice in this repo was exercised with Ramjet before being published.
ADRIAN works without BugLady — the PM conducts the interview and drafts the contract manually. But the framework was designed with BugLady as the requirements engine, and the Request Contract template is what BugLady targets.
ADRIAN was born at Disruptica and is the framework we use internally to build and maintain our products. We open-sourced it because the practices are more valuable shared than hoarded — and because training material, workshops, and developer education derive from this canonical source.
MIT — see LICENSE. Use it, fork it, rename it. If you improve it, consider contributing back via CONTRIBUTING.md.
If you reference ADRIAN in writing, please cite:
Anaya, J. (2026). ADRIAN — AI-Driven, Request-Intelligent, Agentic eNgineering (Version 0.1) [Framework specification]. Disruptica. https://github.com/disruptica/adrian-framework
A machine-readable citation is in CITATION.cff.