Beyang Liu's Frontier Coding Agent

Table of content

Beyang Liu spent a decade building code search infrastructure at Sourcegraph. Then he left to build a coding agent.

The transition wasn’t random. While working on Cody (Sourcegraph’s AI assistant), Liu realized that code search and coding agents represent fundamentally different distribution challenges—and need radically different approaches.

In December 2025, Liu and co-founder Quinn Slack spun off Amp as an independent company focused exclusively on frontier coding agents.

The Spin-Off Logic

Most companies would keep both products under one roof. Liu argued against it:

“Sourcegraph is mission-critical infrastructure software that delivers a consistent experience for enterprises. Amp needs to remain on the frontier, exploring what’s possible with models and tools by building for the devs who want to be a year ahead of everyone else.”

Different audiences. Different distribution. Different companies.

Amp was profitable from the start—its traction justified the split.

Amp’s Philosophy

Amp runs on four principles:

Unconstrained token usage—no artificial limits on model consumption
Always use the best models—Opus 4.5, GPT-5.2, whatever performs best for the task
Raw model power—don’t water down capabilities
Built to evolve—no backwards compatibility, no legacy features

The third point is deliberately aggressive: “If we don’t use and love a feature, we kill it.”

This explains why Amp removed tab completion in January 2026. They decided autocomplete wasn’t part of their vision for the future, so they cut it entirely.

AGENTS.md for Context

Amp reads AGENTS.md files from your project to understand codebase structure, build commands, and conventions. The system loads files hierarchically:

Root AGENTS.md for architecture and build/test commands
Subtree files for subproject-specific guidance
Personal ~/.config/amp/AGENTS.md for device-specific preferences

You can also use granular guidance with YAML front matter:

---
globs:
  - '**/*.ts'
  - '**/*.tsx'
---

Follow these TypeScript conventions:
- Never use the `any` type

The glob-scoped rules only load when Amp reads files matching that pattern. Context stays lean.

“An LLM, a Loop, and Enough Tokens”

Thorsten Ball (one of Amp’s co-founders) published a tutorial called “How to Build an Agent” that strips away the mystique. The core thesis:

“It’s not that hard to build a fully functioning, code-editing agent. It seems like there has to be a secret behind it. There isn’t. It’s an LLM, a loop, and enough tokens.”

Ball demonstrates building a complete code-editing agent in under 400 lines of Go. The rest—what makes Amp “addictive”—is polish and iteration.

The tutorial frames tool use simply: you tell the model what tools exist, it signals when it wants to use them, you execute and return results. That’s the entire mechanism.

Handoff Over Compaction

Amp doesn’t compact conversations when context windows fill up. Instead, they built a “handoff” feature that creates a fresh thread with relevant files and context from the original.

Their reasoning: compacted context loses important details. Starting clean with explicit handoff preserves what matters while dumping accumulated noise.

Amp’s thread sharing lets team members see each other’s agent conversations. The manual encourages it: “Use thread sharing to learn from each other.”

Every Amp interaction can be published with a URL. Browse public threads to see how others approach problems.

The Research Lab Framing

The Amp Inc announcement positions the company as “an independent research lab”:

“Our goal: let software builders harness the full power of artificial intelligence. We believe the way we develop software will change. All of it will change, fundamentally and drastically. Nobody knows exactly how. We intend to find out.”

Then the kicker:

“We believe that shipping is the best way to do that. We don’t want to write papers about the future; we want to put it in your hands.”

Practical Prompting

Amp’s manual includes concrete guidance for prompts:

Be explicit. “Do X” not “Can you do X?”
Keep it short and focused—break large tasks into threads
Don’t make the model guess. Share relevant file paths and commands
Tell the agent how to verify its work: which tests to run, which URLs to check
Abandon cluttered threads. Start fresh when failed attempts pollute context

The first prompt matters most: “It sets the direction for the rest of the conversation.”