Will Bond

Will Bond builds developer tools, from Package Control, the open-source package manager for Sublime Text, to automated code review. At Uber, his work on AI developer experience includes uReview, a collaborative system that routes reviews, filters suggestions, and uses developer feedback to improve their usefulness.

Package Control, Sublime Text, and Python libraries

Bond began by building websites and applications at a web agency, combining programming, design, and client work. In 2011, a deployment plugin he was developing for Sublime Text needed an easier way to deliver updates. He built Package Control to let developers discover, install, and update extensions inside the editor, reducing the manual work of maintaining their tools.

He left his day job in 2014 to concentrate on plugins and software products, then joined Sublime HQ in 2016 as its second engineer. His work on Sublime Text 3 and 4 spanned engineering, visual design, infrastructure, and documentation. In his account of building a high-performance text editor, he emphasizes avoiding unnecessary computation, allocations, and memory use. These choices matter in everyday interactions: an editor that responds quickly lets developers navigate and manipulate code without waiting for the interface.

His open-source cryptography libraries address related problems of distribution and efficiency. oscrypto gives Python applications encryption and certificate functionality through operating-system cryptographic libraries, without requiring users to compile native extensions. asn1crypto parses and serializes structures used in certificates and keys. Delayed parsing, retained encoded data, and lazy field loading let it avoid processing an entire structure when an application needs only a few fields.

Code review at Uber’s scale

At Uber, Bond has worked on editors and IDEs. He co-authored the company's 2025 technical account of uReview, which identified him as a staff engineer in AI Foundations within Development Platform. The system combines specialized reviewers with validation, filtering, and deduplication so that generating a suggestion does not automatically mean publishing it to an engineer.

In their 2026 presentation on uReview, Bond and Ameya Ketkar describe the pressures behind this work. Uber's engineering organization spans hundreds of teams and six language-specific monorepos. They reported that time to first review had increased from three hours in 2024 to nine hours in 2026 as pull-request volume and size grew. An internal platform also needed to support Phabricator during a migration to GitHub and bring consistent review rules into coding-agent workflows.

How uReview delivers useful feedback

• Routing and filtering: uReview accepts requests from human review interfaces and agent loops, then routes them to generators with different performance and cost characteristics. Review depth can depend on a change's risk and complexity, while security and compliance checks need dependable coverage. Post-processing rates, categorizes, filters, and deduplicates the resulting comments to limit the volume engineers receive and prioritize actionable findings.
• Team customization: Team customization is central to the design. General reviewers look for logic bugs within individual files; deeper reviewers incorporate monorepo conventions and examine changes across files. AI linters apply rules with additional context to find systematic issues. Teams can also define custom agents that draw on knowledge bases, previous pull requests, and review skills. Connecting these configurations to Uber's existing ownership model, and keeping them near the code, lets teams maintain their own review knowledge rather than asking a central group to encode every convention.

Learning from developer responses

The collaborators also explain why writing a review skill is easier than operating it well. uReview's evaluation developed from cost tracking, surveys, and support feedback into analysis of developers' replies, whether they addressed comments, and the agents' execution paths. Reply sentiment helps identify unwanted or incorrect suggestions; addressal rates show whether comments prompt changes; tool-call trajectories help reveal inefficient searches and wasted turns. Teams receive these signals so they can revise their rules, while the platform team can tune runtime behavior and cost.

They reported roughly 25,000 comments per week, with about 10% receiving explicit feedback. Around 67% of comments were addressed, rising to almost three quarters for high-severity issues. These are measures of developer response, rather than a direct count of bugs prevented. Against a naive implementation, the presenters reported a 60% cost reduction and approximately 70% improvement in quality and accuracy.

Human judgment in automated review

Their discussion of human and agent review highlights a further challenge. Engineers can become frustrated by many minor comments, whereas agents may readily fix them. Inaccurate feedback can nevertheless send an agent through repeated, contradictory edits. Automated implementation therefore increases the importance of accurate review and raises a question about how to retain the human feedback that helped improve the reviewers.

Bond and Ketkar argue that greater automation should expand human review toward architecture, domain expertise, and product decisions. Their proposal preserves engineers' role in directing the work as agents take on more implementation detail. For uReview, that means building both the machinery for useful automated feedback and the continuing human involvement needed to improve it.