Idan Gazit leads GitHub Next, the research and prototyping organization behind GitHub Copilot and experiments in how software gets built. A designer and engineer who previously worked on Django and Heroku, he focuses on a problem larger than generating code: helping teams collaborate with AI agents without losing context, security, or control.
Gazit was a developer and lead designer on the Django core team, created visual identities for Python and Django conferences, and delivered the DjangoCon US keynote in 2011. At Heroku, he led the design of Heroku App Metrics, became a principal engineer, and founded its Data UX team, which built interfaces for infrastructure products including Postgres, Redis, Kafka, and Cassandra. His personal website describes a career combining software development, design, typography, and the presentation of complex information.
He helped establish GitHub Next as it emerged from GitHub’s Office of the CTO, led developer-experience research, and now leads the wider organization. His projects include Monaspace, an open-source programming-font family whose GitHub Next team included Gazit and whose type design was created by Lettermatic specialists.
- GitHub Agentic Workflows: Gazit wants background automation to handle maintenance requiring judgment, including dependency upgrades, issue triage, failing builds, and repository reporting. Workflows written in natural-language Markdown compile into GitHub Actions; an agent can inspect release notes, update code, verify a build, and propose a pull request. GitHub’s account of continuous AI in practice describes this extension of conventional continuous integration.
- Security boundaries outside the prompt: Gazit argues that instructions alone cannot protect agents against prompt injection. His workflow architecture uses deterministic permissions, network allowlists, mediated secrets, restricted writes, and auditable actions. An upgrade agent can receive permission to inspect package registries and open a single pull request without accessing credentials or contacting arbitrary services.
- Real-time multiplayer software development: His research prototype Ace places teammates, coding agents, shared plans, repository branches, and previews inside cloud-hosted collaborative sessions. Agents can act on decisions already captured in team conversations and collectively edited documents, reducing repeated prompting and misunderstandings. His AI Engineer conference session connects this approach to team alignment as AI accelerates implementation.
- Editable specifications as programming interfaces: Gazit coauthored Extract, Edit, Apply, a technical report describing temporary natural-language specifications generated from existing code. Developers revise the intended behavior, and AI translates those changes back into the codebase. He has also connected programming-language choices to AI-assisted verification, arguing that explicit types help validate generated code while AI can make unfamiliar tools, including shell scripting, more accessible.
