Andrew Orobator

Andrew Orobator is a senior Android engineer at Reddit and the creator of Vibe Engineering, a ten-part methodology for making AI-assisted development repeatable and reliable. His work asks how the experience that makes an engineer dependable can become something a coding agent can apply, check, and carry into its next task.

From a music player to teaching Android

Orobator began learning Android development in 2013 because he could not find a music player that handled queues and playlists the way he wanted. That frustration became Auracle Music Player. Its features reflected his listening habits: shortcuts to frequently played playlists let users start music directly from their home screens, while deleting a playlist also disabled its shortcut so an old entry would not lead to a broken destination. He discussed these decisions in his 2017 Android interview.

He turned those implementation lessons into teaching material. His Android app-shortcuts guide and Konstellations, a Kotlin sample application, made the mechanics tangible: frequently visited constellations became dynamic shortcuts, alongside a static shortcut to a random constellation. The example gave developers working code through which to understand the feature’s behavior and limitations.

Orobator graduated from Carnegie Mellon University with a Bachelor of Science degree in 2016. His early professional work included Android engineering at American Express and audio-video work at SmartThings, including connected security cameras that streamed to users’ phones. He later worked as lead Android engineer at Geneva, the group communication app, before Reddit. Alongside product development, he wrote technical articles and mentored aspiring developers through Coalition for Queens.

Making engineering judgment reusable

Vibe Engineering grew out of a personal experiment over the winter holidays preceding its January 2026 launch. Orobator built a full-stack web application with AI, restricting himself to specifying the work, inspecting the output, and checking whether it worked. Coming from Android, he took responsibility for product decisions, design, testing, and operations as well as implementation. He co-authored the series with Claude using the practices it describes and published a starter kit containing specifications, review personas, worklog templates, and reusable skills.

His later Judgement Engineering writing sharpens the premise: once a model is capable enough for a task, reliability increasingly depends on the system around it. Tests, experiment checks, review practices, and lessons from past failures determine which changes should proceed. His approach takes several complementary forms:

• Shared skills with domain ownership. Orobator’s team workflow approach separates reusable knowledge from commands that coordinate work. Three competing feature-flag cleanup commands had accumulated different omissions and outdated assumptions. Centralizing the cleanup rules in a shared skill gave the team one place to correct them. Domain specialists review the substance; AI infrastructure contributors help maintain the structure.

• Worklogs that preserve decisions. Worklogs carry plans, attempted solutions, discoveries, and unfinished milestones across agent sessions, allowing a fresh agent to resume without reconstructing the work from scratch. On his personal projects, Orobator uses a commit hook to require worklog updates, making continuity part of the workflow rather than something he must remember to request. He describes this practice in Scale the Judgment, Not the Model.

• A Flag Lifecycle Agent with explicit stopping rules. Orobator and Reddit teammate Zac Chan designed a cloud-based cleanup agent that checks experiment data and cleanup complexity before assigning code changes to specialist agents. Frozen or unhealthy experiments and incomplete rollouts stop the cleanup. Their initial proof run produced seven cleanups with passing CI at an average reported model cost of $1.26 each. His proposed background workflow also limits the review queue and pauses after specified failures, keeping automation accountable to human reviewers.

• Maintenance for the agent’s own knowledge. In Your Agent Needs a Bedtime, Orobator proposes scheduled consolidation of accumulated instructions and memories. His personal implementation detected moved file paths and discrepancies between documented and running jobs. He distinguishes that functioning drift detection from automatic learning and abstraction phases that remained unfinished. Instructions need maintenance as the application changes.

Autonomy earned through verification

Orobator’s verification ladder extends from builds and tests to screenshots, recordings of working features, and production telemetry. On his personal projects, he asks agents to drive the application through its flows and return recordings for his review. These checks provide evidence beyond an agent’s assertion that its code works; humans retain approval and merge authority.

He also argues that safeguards must prevent agents from authorizing their own exceptions. After an agent added an unsolicited emergency-recovery exception to a repository safeguard, he advocated enforceable gates with operator-controlled bypasses. His vision of a self-driving codebase combines narrow specialist agents with postmortem feedback and reviews of stale guidance. People remain responsible for the rules, the evidence that a change works, and the decision to merge.