Why an AI Assistant Belongs in the Terminal
Developers spend their day doing work that never touches an IDE: SSHing into servers, debugging containers, managing CI/CD pipelines, writing deployment scripts. An IDE-only assistant covers maybe half that surface. The terminal covers all of it.
“the terminal transcends all the different applications” — Ryan Hecht
Hecht’s argument is that agentic AI in the terminal removes the two biggest friction points: hunting for the right flags and reading man pages. You get fine-grained control, composable scripting, and the full breadth of open-source tooling, with Copilot filling the gaps you’d otherwise pause to look up.
Getting Started in an Unfamiliar Codebase
Hecht demoed taking over on-call for a colleague’s feedback form app without knowing the language, framework, or whether the repo was even cloned. He showed the unfamiliar repo onboarding demo (03:21) live: Copilot read the README, detected a package.json, ran npm install, then diagnosed a port 3000 conflict by running lsof with the correct flags. The whole sequence ran on natural-language prompts, no documentation consulted.
“it’s way easier to just defer that to Copilot.” — Ryan Hecht
Copilot also accepted an image file showing a misaligned CSS button and located the exact file to change, without the developer opening a single source file manually.
Custom Agents and Inline Compliance Checks
GitHub Copilot CLI shipped custom agents on the day of the talk. A team encodes its domain rules and the MCP tools to call into a named agent. In the accessibility agent review demo (08:34), Hecht invoked his team’s “accessibility reviewer” agent, which scanned the pending CSS changes, found a critical issue, and applied a fix with no manual rule lookup.
The agent also queried open GitHub issues via the bundled GitHub MCP Server and surfaced a separate mobile-sizing bug, adding it to the requirement list before Hecht moved on.
Headless Mode and Delegating Work Asynchronously
The delegate command is a new addition that stages current changes, commits them to a fresh branch, opens a PR, and dispatches Copilot Coding Agent to finish the remaining work. Hecht showed the delegate to Copilot Coding Agent demo (10:33): he passed the open issue number and a Playwright MCP Server instruction, then moved on while the agent worked in the background.
The -p flag enables headless, scriptable execution. Combined with permission flags that restrict which directories and tools Copilot can touch, it fits into GitHub Actions workflows or any automation that needs AI reasoning without interactive prompts.
Notable Quotes
the terminal transcends all the different applications Ryan Hecht · ▶ 1:34
it’s way easier to just defer that to Copilot. Ryan Hecht · ▶ 5:21
we’ve shipped a new release every single working day. Ryan Hecht · ▶ 13:21
Key Takeaways
- Copilot CLI runs as an npm module on Mac, Linux, and Windows, installable in one command.
- The delegate command hands unfinished work to Copilot Coding Agent, which opens a PR and works asynchronously.
- Custom agents encode team-specific rules and MCP tools, so compliance checks run before any commit.