The breadth of GitHub Universe sessions spans concerns ranging from how to validate code written by AI agents to understanding the trust relationships embedded in every dependency installation. The conference program addresses three core questions: How can we verify that an agent's code actually works? What should agents be allowed to remember? And what are the security implications of the packages we pull into our projects?
With more sessions than any attendee can fit into two days, the challenge becomes prioritizing talks that address immediate technical challenges. Agent memory, code evaluation, and permission models emerge as high-priority topics, alongside practical concerns like JavaScript tooling modernization and building applications where network connectivity cannot be assumed.
Dependency security and npm internals
Karen Li and Leo Balter from GitHub will examine what happens when developers run npm install, tracing how dependencies move through the systems that publish and protect them. The session will explore gaps in what npm audit can detect, and how package provenance and OpenID Connect mechanisms address those blind spots.
Agent memory and context management
GitHub researchers Cooper Nederhood and Alejandro Carderera will present findings from a benchmark built on real pull request sequences, revealing that accumulated context can degrade agent performance. Their work on agent memory—including unreleased features—will address what information to withhold from agents, not just what to provide.
Christopher Harrison from GitHub will discuss treating AI context as infrastructure, examining how to distribute consistent context across teams as tooling setups grow in complexity. He will break down which tools serve which purposes and how to scale context distribution reliably.
Nick Taylor from Pomerium will demonstrate fine-grained authorization for hosted Model Context Protocol (MCP) servers using an identity-aware proxy. His session will show how to enforce rules like "open a pull request, but do not merge" through per-identity authorization without modifying upstream servers.
Code verification and evaluation
Walker Chabbott from GitHub and Julia Kasper from Microsoft will explore how Copilot evaluates models in production, covering what metrics the team tracks and what they have stopped measuring. This Sandbox Session will examine what makes an evaluation useful when a model can score well on benchmarks yet disappoint developers in real use.
Jeff An from Momentic will examine how agents verify AI-generated code by investigating applications, reproducing unexpected behavior, and distinguishing product bugs from broken tests or infrastructure failures. The session will focus on deterministic controls that constrain agent actions during investigation.
Achin Gupta from Intuit and Divya Mahajan from Amazon will present an architecture where deterministic code handles signal collection, topology traversal, correlation, and scoring, while the language model narrates the evidence. This separation of responsibilities clarifies how to build AI systems that investigate rather than hallucinate.
Supply chain security and tooling
Steve Glass and Greg Ose from GitHub will map supply chain attack techniques to GitHub Actions controls, covering the ecosystem, workflow attack surface, and runner infrastructure. The session will help attendees understand where to position defenses in pipelines trusted with releases and credentials.
Alexander Lichter from VoidZero will present Vite+, an open source CLI for managing the front-end toolchain. The session will include a real migration example, showing what Vite+ replaces today, what requires separate management, and what remains on the roadmap. Bundling, testing, linting, formatting, and runtime management will be covered.
Building for low-connectivity environments
Alex Junior Antwi from Braveon AI will share lessons from building CarbonSight for low-connectivity communities in Ghana. The session will address the long list of concerns when designing for less-than-ideal connectivity scenarios and will include an offline demo.
Building reusable agent skills
A practical session titled "Build once, run on any agent: a practical guide to agent skills" will bring together Shishir Tewari from Procore and an attendee with experience building skills for recurring work. Tewari will contribute insights from building skills for a production data engineering team, exploring when a workflow belongs in a reusable skill and what makes a skill useful across both work and personal projects.
GitHub Universe 2026 runs October 28–29. In-person and virtual passes remain available.