AI Coding Assistants

Today's Stories

1. 1

   Enterprise AI evaluation shifts from single-conversation scoring to cohort comparison

   Leaders from LangChain, Conviva, and CoreWeave said at VB Transform 2026 that enterprises are moving away from scoring individual AI agent conversations in isolation—which can appear flawless but mask broken products—toward comparing groups of users against a baseline. The industry is also adopting cheaper, narrower judge models instead of larger ones. A single conversation that looks perfect on its own does not tell you whether your AI product actually works for real users. This shift means enterprises can now catch systemic problems that individual-trace scoring misses, improving how well AI agents perform in production.

   The tension between automated judging (whether by AI agent or large language model) and human review remains unresolved; Zhang noted that automated methods are scalable but hard to ground in real outcomes, while human review is thorough but does not scale.

2. 2

   Adobe's Indigo camera app adds generative AI editing tools

   Adobe has updated its Indigo camera app with an "AI Playground" suite of generative AI tools — including style filters, object removal, photo guidance suggestions, and a custom edit prompt tool. The app now uses Google's Nano Banana AI model for these features. Free access is being tested with a small percentage of Indigo users over the next few weeks, with an opt-out button available. Indigo launched last year as a camera app focused on computational photography and manual controls to replicate SLR-quality iPhone photos. Adding generative AI editing directly inside the app transforms it from a pure capture tool into an integrated mobile camera-and-editing experience. Adobe says it is experimenting to learn how people use AI-powered editing and may eventually offer a paid version if the feature proves popular.

   Adobe is working toward adding C2PA Content Credentials — a metadata standard for identifying AI-manipulated content — to images edited by Indigo's tools. The company notes that Nano Banana already applies Google's invisible SynthID watermark. Adobe fellow Marc Levoy emphasized that responsibility in use remains the user's responsibility, not just the tool's.

3. 3

   Chrome extension adds AI side panel to LeetCode coding problems

   A Chrome extension called LeetCopilot has been created that integrates an AI side panel directly into LeetCode, the popular coding interview practice platform. LeetCode users preparing for technical interviews can now access AI assistance without leaving the problem page, potentially reducing context-switching and making practice sessions more efficient.

   The extension is available at leetcopilot.dev; adoption and user feedback will determine whether this workflow integration model gains traction among coding interview candidates.

4. 4

   Augment Code's semantic context engine offers alternative to Claude Code's lean AI coding harness

   Augment Code's VP of Engineering Vinay Perneti outlined the company's approach to AI-assisted coding, which pre-indexes code repositories using embeddings and a retrieval model to surface contextually relevant code—a contrast to Claude Code's grep-based, lighter approach that avoids structured context by default. In benchmarks using the same model, Augment Code achieved 33 percent greater token efficiency than Claude Code while maintaining similar accuracy, suggesting that richer context handling can reduce wasted token spending on code exploration. For development teams, especially those with large private codebases that AI models have never encountered, this design choice affects both the quality of outputs and the cost of agentic workflows.

   The debate reflects a fundamental tension in harness design: as models improve exponentially, whether to bet on the models' raw intelligence to find their own way (Claude Code's lean bet) or invest in engineering infrastructure to guide them toward the code they need (Augment Code's semantic approach). Perneti argues the answer involves both intelligence and context, framed as a systems engineering problem of balancing token efficiency against outcome quality.