AI
From Toy to Tool: Closing the Reliability Gap
GitHub Next is GitHub’s long-bets team. Copilot, Copilot Workspace, and Spark all started there. Getting an API key and building something that works half the time takes an afternoon. Making it reliable enough to ship is the actual work.
Gazit’s framing: interactive agents already help with blank-canvas moments, but the bigger opportunity is background work. Tasks that always need doing but don’t require a human present. That’s what continuous AI is built for.
Continuous AI as the Agentic Evolution of CI
Continuous AI extends continuous integration from code-expressed rules to natural-language rules. GitHub Actions already runs 60 million jobs a day. Describing a rule in plain English is now enough to automate it.
The docstring checker is the clearest example. A traditional linter flags a missing docstring. It can’t detect that a function marked “returns a copy” is actually mutating the input. An agentic workflow catches that mismatch and opens a pull request with corrected code. You write the workflow as a markdown file: a front matter block, then a natural-language prompt.
Live Demo: Building an Agentic Workflow in Minutes
Pelayo (Microsoft Research) built the daily news report built live in VSCode (06:45): a plain-English prompt in Copilot, a few clarifying questions about which events to track, and Copilot CLI compiled the result into a GitHub Actions YAML file that posts a discussion every weekday morning.
The same session showed a dependency watcher catching an undocumented flag (10:21) in a CLI the team depends on. The agent installed the tool, read the help output, cross-referenced previous days, and filed the finding. Pelayo typed the entire workflow on a phone to make the point that a laptop is not required.
Continuous AI Across the Full SDLC
A 2-character config change triggers an agentic workflow that adds Korean translations in a pull request and keeps them current on every future commit. GitHub’s accessibility team shipped a scanner that finds text contrast issues and opens PRs to fix them. Three agent personas play a hackathon platformer at hyper speed in the three-persona game playtesting catching broken difficulty balance (16:07): they found that lava blocks capped even the high-skill player below a score of 23. The 45-day test coverage run on fs-math (18:42) took that library from 5% coverage to 100%, writing 1,400 tests for $80.
Safe Outputs and the Threat Model
Every agentic workflow ships with four security controls. Prompt protection keeps the agent listening to the workflow author, not injected instructions in user-filed issues. Fine-grained MCP permissions cap what each server is allowed to do. A web firewall prevents agents from browsing arbitrary sites that could carry malicious instructions. Safe outputs are the last line: agents are read-only by default, and each workflow declares an explicit allowlist of what it can create. The performance improvement bot, for example, can write a discussion, post a comment, and open a pull request. Nothing else gets through.
Notable Quotes
The only way to know is to make. Idan Gazit · ▶ 1:43
Backlog fairies are real. Idan Gazit · ▶ 19:58
It went from 5% test coverage on this library to effectively 100% coverage. Idan Gazit · ▶ 18:49
Key Takeaways
- Agentic workflows compile plain-English prompts into GitHub Actions YAML, no boilerplate required.
- A 45-day continuous AI run took fs-math from 5% to 100% test coverage for $80 in tokens.
- Safe outputs make agents read-only by default, limiting what they can create to an explicit allowlist per workflow.