Amp Code: The Terminal-First AI Agent That's Living One Year in the Future
Why Sourcegraph's experimental agent uses sub-agents, custom tools, and ads to push the frontier of AI coding. Discover the "For Loop" philosophy, the 80% success Oracle, and an industry-first ad-supported inference model.
We really want to be the agent research lab that's sort of like living one year in the future and figuring out how this all kind of pans out.
— Beyang Liu, Co-founder & CTO, Sourcegraph (00:01:10)
Dual UI for different workflows
Finder, Oracle, Librarian, Kraken
Industry-first inference sponsorship
The "Opinionated Agent" Philosophy
Amp isn't trying to be for everyone. They're targeting "the small percentage of people who want to live a little bit in the future" — and embracing the weirdness of agents writing code.
We're the bunch of weirdos downstairs at the booth with the weird pied piper dude on the floating golden fish. I think that kind of captures the ethos of what we're trying to do.Watch (00:00:45)
Not for Everyone
"We're really trying to target the small percentage of people who want to live a little bit in the future."
No Influencer Marketing
"We're not about hype. We don't do any sort of like paid developer influencer marketing."
Research Lab, Not Product
Amp is an experimental project by Sourcegraph, separate from their main Cody AI product.
Architecture: The "For Loop" Philosophy
All an agent is, fundamentally, is "a for loop with tool calls and a model in the middle." From that simple primitive, Amp makes three contrarian architectural decisions.
Well, all an agent is... is it's a for loop uh with tool calls and a model uh in the middle.Watch (00:04:05)
Why We Chose Custom Tools Over MCP
While most of the industry rallies around Anthropic's Model Context Protocol (MCP), Amp deliberately chose custom tools.
"The more tools that you add into the context window, the more things that the agent has to choose from. And if the tools aren't relevant to the task at hand, it ends up getting confused."
The Problem with MCP:
- • MCP servers are general-purpose (creator doesn't know your agent's needs)
- • Tool descriptions can't be optimized for specific workflows
- • Context confusion from too many tools in the prompt
"This is maybe, you know, less controversial now than it was back in April."
Watch (00:04:15)Sub-Agents: The Subroutine of Agent Architecture
Sub-agents are how you factor out the context window used for a subtask into a separate context window — like subroutine calls in programming.
"Sub-agents are the analog to subroutine calls in regular programming languages. This is how you can factor out the context window used for a subtask into a separate context window."
The Context Problem:
- • "Context exhaustion": Agent reads too many files, no tokens left for editing
- • "Doom loop mode": Insufficient initial context leads to endless retries
- • Solution: Sub-agents have separate contexts, return only relevant results
Two Agents, Not N Models
Amp rejects the "model selector" pattern used by Cursor, Windsurf, and others. Why? The paradox of choice.
"If you support N models, you can only lightly customize each. You're never really optimizing for what any one given model can do."
Amp's Two-Agent Architecture:
- • Smart Agent: Slow but capable, has all sub-agents
- • Rush Agent: Fast, for quick in-the-loop edits
- • Recently switched Smart Agent to Gemini 3.0 (two days before talk)
"The problem with choice is that there's also a paradox of choice. The more choices that you have, the more cognitive burden it is to choose."
Watch (00:09:13)The Four Sub-Agents
Each sub-agent specializes in a specific task with its own context window, model selection, and tool set. Together, they solve the context bottleneck.
Finder
Codebase search sub-agent that uses a small, quick model with limited tools optimized for discovering relevant context.
"We've ended up at the point now where we're using a relatively small and quick model to drive a limited tool set that we found really is optimal for quickly discovering relevant context within the codebase."
Oracle
Deep reasoning sub-agent for debugging tricky problems and nuanced planning. The hero sub-agent with remarkable success.
"Four or five out of five times it just magically finds the underlying issue."
Librarian
Fetches context from external libraries and frameworks that your codebase depends on.
Kraken
Large-scale refactoring sub-agent that writes code mods for automated refactors.
Why Terminal AND Editor?
Most AI coding tools are editor-only. Amp does both terminal and editor interfaces because different workflows need different modalities.
Async Task Spinoff (Smart Agent)
"You kind of like spin off a task and have it run and then review the code when it's finished asynchronously."
In-the-Loop Babysitting (Rush Agent)
"You want to be in the loop, quickly having the agent make edits while you quickly review them one by one. Kind of like babysitting the agent in the inner loop."
The Terminal-First Approach
Custom TUI Framework
- • Built from scratch for optimal performance
- • Streams diffs and CLI commands in real-time
- • Shows right amount of information without overwhelming
Editor Integrations
- • Connects to Emacs, Neovim, JetBrains for diagnostics
- • Ghostty integration — core contributor built AMP's TUI
- • Graceful degradation: works in default Mac OS terminal
The Real Bottleneck: Code Review Speed
"I spend most of my time effectively doing code review now. That's the thing that constrains me from fully paralleling 2-3x the number of agents."
The solution: Custom diff viewer with "tour of the change" that guides you through which files to read first.
Industry-First: Ad-Supported AI Coding
One of the biggest barriers to using AI coding agents is cost. Amp's solution? "Ship ads in your terminal" — an industry-first experiment in sponsored inference.
The Cost Barrier
"If you go to like college campuses and talk to students, the actual number of people who have used a coding agent is actually much smaller than I would have thought given young people's propensity to adopt new technology. A lot of this is cost."
Rush models are getting cheaper and cheaper but not yet free. Students and hobbyists can't afford $20/month subscriptions.
Watch (00:14:41)The Crazy Idea
"Someone had the crazy idea on our team like hey you know what we could do we could ship ads in your terminal. And at first it was like nah that'll never work. But the more and more we thought about it... yeah maybe."
As inference costs declined, ads became viable. "At first it was like nah that'll never work. But the more and more we thought about it... yeah maybe."
Watch (00:14:59)Mini Ad Network in AMP
"So we actually shipped a mini ad network that delivers ads for other developer tools in AMP in the terminal and in the editor."
Who's Actually Using Amp?
Amp doesn't do paid developer influencer marketing. Here are two notable users who genuinely use the product:
Mitchell Hashimoto
Founder of HashiCorp (Vagrant, Terraform, Consul, Vault). Creator of Ghostty terminal emulator.
"Using AMP to drive a lot of the changes he makes to Ghostty."
⚠️ Name Correction: Transcript shows "Hamill Hussein" — actual name is Hamel Husain
Hamel Husain
AI/ML engineer, specialist in AI evals and LLM applications. Leading authority on AI evaluation.
"As of a couple weeks ago he was saying that AMP was his favorite coding agent."
The Community of Builders
Run by Ryan Carson (former Treehouse founder, taught 1M+ people to code), the community is for "builders working with AI agents — using Amp is NOT a requirement."
Regular interviews, in-person events, philosophical discussions about AI nature, and technical deep-dives on agentic workflows.
"It's all about kind of like tapping into that sense of awe and wonder with a peer group that is also leaning into that sense of strangeness and experimentation."
Key Takeaways for AI Engineers
Sub-Agents Beat Model Switchers
Specialized sub-agents (Finder, Oracle, Librarian, Kraken) with separate context windows outperform single agents with model selection.
Custom Tools > Generic Protocols
MCP servers cause context confusion. Custom tools tuned for your agent's specific workflow enable better feedback loops.
Terminal + Editor Covers All Modalities
Some workflows need async spinoff (terminal), others need in-the-loop babysitting (editor). One size doesn't fit all.
Context Exhaustion is the Real Bottleneck
The "doom loop mode" (insufficient context) vs. premature stopping (too much context). Sub-agents solve this by factoring out context usage.
Ad-Supported Inference Enables Access
Students can't afford expensive agents. Subtle ads in terminal/editor sponsor free Rush Agent access.
Two-Agent Model Simplifies Choice
Smart (slow, capable) vs. Rush (fast, in-the-loop). Avoids paradox of choice from N model options.
Companies & Technologies Mentioned
Gemini 3.0
Mentioned as released "two days ago" in Oct 2024 talk
Model Context Protocol
Anthropic's open protocol for AI tool integration
Amp chose custom tools instead
modelcontextprotocol.ioSource Video
Amp Code: Next Generation AI Coding
Beyang Liu, Sourcegraph / Amp • AI Engineer Summit 2024
Research Note: All quotes in this report are timestamped and link to exact moments in the video for validation. This analysis was conducted using multi-agent transcript analysis with transcript-analyzer, highlight-extractor, fact-checker, and ui-ux-designer agents.
Fact-Check Accuracy: 9.5/10 — Key corrections: "Hamill Hussein" → "Hamel Husain", Gemini 3.0 timeline discrepancy noted.
Technologies Mentioned: Sub-agents, Model Context Protocol (MCP), Terminal UI, Custom tools, Context window management, Ad-supported inference, Gemini 3.0, Ghostty, JetBrains, Neovim, Emacs