# We Will Fix Everything

> A guide by Andrius Bartulis about software engineering in the age of AI. AI gives us the tools to build substantially better software. Humans own product decisions and the quality bar; agents help explore, implement, challenge and verify the work.

This site presents the author's proposed working practices and opinions. It does not claim to run the automated infrastructure it describes.

## Read

- [Guide](https://wewillfixeverything.com/): Canonical HTML, complete without JavaScript.
- [Full plain-text edition](https://wewillfixeverything.com/manifesto.txt): The full argument, working loop, release standards and planning targets.

## Interactive lab

- [SDLC simulation](https://wewillfixeverything.com/simulation.html): Four presets compare traditional development, AI with unchanged PR approval, YOLO and guided AI. A synchronized 180-day comparison shows satisfaction, quality and unmet customer needs, alongside released-versus-useful output bars. All four approaches advance together in about 15 seconds; selecting one highlights it and opens its optional workflow without resetting time. Every scenario shares an initial backlog, uneven arrivals and growing demand. Defects persist in use during repairs; satisfaction responds to useful delivery, defects and delays. YOLO adds parallel capacity and speculative scope, prioritizing features over repairs. It can deliver more releases but less useful product in the long run. All numerical rates and customer indices are teaching assumptions, not research-derived estimates.

## Principles

- [Raise the quality bar](https://wewillfixeverything.com/#commitment): Use AI for deeper exploration, better design and stronger verification. Quality is a moat alongside reputation, service and customer relationships.
- [Own the product and the decisions](https://wewillfixeverything.com/#features): Human judgment, taste and accountability throughout design, implementation, product review and iteration.
- [Demand evidence](https://wewillfixeverything.com/#multiplier): Real user journeys, independent checks and measured results. Human attention goes to consequential decisions, with explicit authority for routine changes.
- [Shorten the path to excellent software](https://wewillfixeverything.com/#ownership): Reduce waiting and rework, use reproducible environments and cut scope while preserving quality. A one-to-two-week MVP planning target; readiness determines release.
- [Keep improving after release](https://wewillfixeverything.com/#quality): Learn from real use and feedback, continually challenge staging, monitor production and verify recovery.

## Extra reading

- [Dark factories and the limits of automation](https://wewillfixeverything.com/#dark-factories): The author argues for continued human product judgment, drawing on research about underspecification, LLM judge biases and reward hacking. The studies establish specific limitations, not inevitable failure of every autonomous pipeline.

## Author and references

- [Andrius Bartulis on X](https://x.com/AndriusBartulis)
- [Andrius Bartulis on LinkedIn](https://www.linkedin.com/in/andrius256/)
- [Toyota Production System](https://en.wikipedia.org/wiki/Toyota_Production_System)
- [The Goal](https://en.wikipedia.org/wiki/The_Goal_(novel))
- [StrongDM software factory approach](https://factory.strongdm.ai/)
- [D’Amour et al., Underspecification (JMLR, 2022)](https://www.jmlr.org/papers/v23/20-1335.html)
- [Zheng et al., Judging LLM-as-a-Judge (NeurIPS, 2023)](https://arxiv.org/abs/2306.05685)
- [MacDiarmid et al., Natural Emergent Misalignment from Reward Hacking (2025 preprint)](https://arxiv.org/abs/2511.18397)