Chris Kelly leads product for Cosmos at Augment Code, developing AI coding tools for teams managing complex production software. Previously FireHydrant’s chief technology officer, he brings experience in developer advocacy, open-source communities, incident response, and engineering productivity to a central question: how can AI help engineers change software safely?
Kelly was a developer evangelist at New Relic by 2013 before joining GitHub, where he helped advance corporate participation in open source. In 2014, he authored the announcement introducing the TODO Group, a collaboration among companies including GitHub, Google, Facebook, and Twitter. He subsequently worked at Salesforce before becoming chief technology officer at FireHydrant, where he championed service catalogs and clearly assigned incident-response roles to help teams locate failures and coordinate recovery.
At Augment Code, Kelly has focused on AI assistants for large, complicated codebases and agents operating across the software-development lifecycle. His priorities include:
- Production software demands more than code generation. Every generated line creates maintenance, security, debugging, and operational obligations. Kelly’s critique of vibe coding distinguishes plausible implementations from software that remains reliable under real production conditions.
- Codebase context enables safer agents. Internal APIs, architectural decisions, dependencies, documented standards, reproducible environments, fast tests, and clearly defined tasks provide the grounding both engineers and agents need to make responsible changes.
- Code review becomes a critical engineering skill. As agents increase output, engineers must evaluate architectural consequences and distinguish genuinely better implementations from merely different ones. Kelly also challenges review interfaces that organize changes alphabetically instead of around system behavior.
- Agents should work beyond the editor. His vision of software engineering beyond the IDE places agents wherever work begins, including bug reports, production alerts, and failing tests. His define-create-refine workflow pairs explicit planning and agent implementation with human inspection. In 2026, he coauthored an account of rebuilding the Auggie CLI agent harness, addressing context retrieval, tool orchestration, session management, and efficiency.
