Beyang Liu's Agentic Coding Setup

Table of content

Beyang Liu co-founded Sourcegraph to bring Google-quality code search to every developer. Now he’s building Amp, an agentic coding tool that’s reshaping how he writes software.

The Inner Loop Problem

Before AI agents, Beyang thought deeply about developer productivity through a systems lens. His core insight: there are two nested loops in software development.

The outer loop is the SDLC—sprints, releases, deployments. Everyone talks about this.

The inner loop is what actually produces code: read → write → run → repeat. This happens dozens of times per day when you’re in the thick of understanding and modifying code. Most productivity discussions ignore it entirely.

“It’s important to talk about the inner loop. It’s the heart of software creation. If you don’t talk about it, then your organization will treat it as if it doesn’t matter.”

The goal is reaching flow state—that golden condition where you’ve paged in all necessary context and can code at the speed of thought. Interruptions destroy it. Both external (meetings, Slack) and internal (needing to understand a library, resolve a blocking issue).

How Agentic Coding Changes Everything

With agents, the inner loop transforms. Beyang now generates more than 90% of his code through Amp rather than typing it himself.

His workflow has two modes:

Writing mode: Detailed prompts to the agent in a side panel
Review mode: Examining diffs, staging changes, fixing issues

The actual editor pane is now secondary. The primary views are the agent panel and the diff view. He alternates between them constantly.

“Agents fetch their own context through the use of tools. As a result, we no longer need much of the UI developed for the chat LLM era where the human was deeply involved in curating context and steering the LLM at every step.”

The Practical Setup

Beyang uses Amp two ways:

CLI in tmux for parallelizing lightweight tasks across multiple worktrees
VS Code extension for larger, complex features

His keybindings map to the home row for rapid switching:

⌘⌥H — Diff view
⌘⌥J — Diagnostics panel
⌘⌥K — Debug console
⌘⌥L — Terminal
⌘⌥; — Amp agent

One VS Code trick: in the diff view, code intelligence still works on the green side. You can jump to definitions and find references without leaving the review.

Thread Hygiene

Beyang prefers creating new threads for subsequent tasks. LLMs perform better when less of the context window is used.

The exception: when the next task is closely related to what you just did and benefits from existing thread context.

What Remains Valuable

Despite the agent generating most code, the editor isn’t dead:

“I still rely on the editor for the rich diff view and diagnostics, which enable my human brain to review and grok the generated code.”

He installs Amp into vanilla VS Code with Copilot disabled. Most AI IDE features (RAG chat panels, context curation UI) become unnecessary noise when the agent handles its own context.

Developer Productivity as Developer Hertz

Before agents, Beyang proposed a unit for measuring developer productivity: iteration frequency—how fast you cycle the inner loop. Not lines of code. Not commits. Just: how quickly can you read-write-run?

Agents accelerate this cycle dramatically. But the human review step remains essential. The skill shifts from writing code to evaluating code.