Amp Code After the Neo Rebuild: We Tested the New CLI

On May 6, Sourcegraph rebuilt the Amp Code CLI from the ground up. The numbers they published: CPU usage dropped 79%, memory dropped 70%. We installed the Neo build, ran it on 5,000-message threads, and watched it handle context loads that would have choked the old version.

Amp Code is a frontier coding agent—part of Sourcegraph’s broader suite, spun out with its own identity and pricing. It’s different from Claude Code because it’s team-focused, built on pay-as-you-go pricing, and ships with specialized sub-agents (Oracle for research, Librarian for codebase understanding). The Neo rebuild is the inflection point: this is the version where Amp stops feeling like an experiment and starts feeling like a tool that scales with your threads.

The Neo Rebuild in Numbers

The rebuilding started from a fundamental architectural problem: the old Amp CLI choked under load. When a thread accumulated 5,000+ messages—which happens faster than you’d expect in serious coding workflows—the system would burn CPU, leak memory, and slow to unusable speeds.

Sourcegraph’s response was not a patch. It was a rewrite.

The Neo rebuild shipped with a “remote-controllable, compaction-first, plugin-powered” architecture. In stress tests on a 5,000-message thread, the new version cut CPU mean usage from 84.1% down to 17.4%—that’s the 79% drop. Idle memory fell from 1,814 MB to 540 MB. Peak CPU usage dropped from 86.3% to 25.8%.

Translation: you can now keep threads running longer, iterate more quickly, and actually afford the token costs when you’re working at scale.

What Amp Actually Is (Now That It’s Spun Out)

Amp started inside Sourcegraph but has evolved into its own tool. It’s a CLI agent that runs coding tasks autonomously, built on frontier models (Claude Opus by default, with flexibility for other models).

The two sub-agents that matter:

• Oracle: runs research and planning phases, pulls context, and briefs you before the agent commits code changes
• Librarian: digs through your codebase and synthesizes what the agent needs to know

Both are designed to let teams keep coding agents on a leash—you don’t just let them hallucinate-code and pray. You ask them to understand the system first, then write.

Sourcegraph’s product page also mentions AGENT.md files, which let you define team coding policies (lint rules, testing standards, commit message formats) right in your repo. It’s a small idea with a big upside: when everyone’s using the same AI coding agent, you want to bake consistency into how it thinks.

How It Feels After the Rebuild

Stephanie Jarmak, a Solutions Engineer at Sourcegraph, spent four months and 6,000+ threads using Amp. She structures her work in short, focused threads (1-3 related tasks per thread) and monitors context window usage to avoid token-cost blowups. Her workflow—Oracle for planning, code generation, then human review—maps to how teams actually work.

What stands out from her report: Amp “feels way more agentic” than point solutions. It doesn’t just autocomplete code; it reasons about your architecture, asks clarifying questions, and flags risky changes. The Neo rebuild made that experience actually usable over long sessions.

Is it perfect? No. Jarmak notes that verification is still your job—the agent can miss edge cases, and technical debt sneaks in if you aren’t reviewing the output. But the feedback loop is fast enough that you can catch it before it lands on main.

Amp vs Claude Code

Both use Claude Opus 4.7. Both handle large token windows. The differentiation lives in the team layer and the pricing model.

Claude Code (Anthropic’s direct integration) is simpler—you get Claude in your IDE, you pay per token, and you’re done. It’s great for solo devs or small teams that want minimal setup.

Our Claude Code review covers the token-limit mechanics in depth. Amp is team-optimized. The Oracle and Librarian sub-agents add structure. The AGENT.md policy layer means you’re not reinventing the same code standards in prompts every time. And the pay-as-you-go model (no subscription, no per-seat licensing) scales differently as your team grows.

If you’re shipping alone, Claude Code probably wins on simplicity. If you’re coordinating a team of 3+ engineers and want the agent to respect your architectural decisions, Amp’s team features pay for themselves.

Pricing: The Part Nobody Likes Talking About

Amp uses a pay-as-you-go model with “no markup for individuals.” Translation: you’re paying Claude Opus token rates directly, with no platform fee on top.

That’s honest. It’s also dangerous if you’re not watching your usage.

A 5,000-message thread with Opus 4.7 is not cheap. Neither is running Oracle and Librarian searches across a massive codebase every iteration. Sourcegraph advertises “frontier model usage”—meaning you get the best models available, and you pay what those models cost.

If you’re heavy on tokens (and teams using agents usually are), budget accordingly. There’s no UI to set spending caps, so it’s on you to monitor. The Neo rebuild at least makes it possible to run 6,000 threads without accidentally burning money on wasted CPU cycles.

What Amp Doesn’t Do

Amp is for code. If your team also ships writing—documentation, marketing copy, changelog posts—that’s a separate problem. Jasper handles the writing layer. Amp doesn’t.

Amp also doesn’t support fully private deployments or bring-your-own-model setups. If your enterprise requires VPC isolation or the ability to run your own fine-tuned models, Cursor still has advantages.

And threads are stored on Sourcegraph’s servers by default, not locally. If your contracts demand on-prem everything, you’ll hit walls.

The version before Neo had other gaps—custom model selection was missing, data-usage policies were unclear—but those feel like surface fixes. The deeper architectural limits are real.

Where Amp Fits in Your Stack

Solo dev or small team (2-3 engineers): Amp is overkill unless you care deeply about code consistency policies. Claude Code or GitHub Copilot Agent is probably enough.

Growing team (4-10 engineers): This is where Amp shines. The Oracle-Librarian workflow prevents hallucinations. The AGENT.md layer keeps everyone’s code standards aligned. The pay-as-you-go model is cleaner than per-seat licensing.

Enterprise (10+ engineers): Amp still requires server-side storage of conversations and works best inside Sourcegraph’s hosted environment. If your compliance team demands full on-prem, Amp’s positioning doesn’t fit yet.

Verdict

The Neo rebuild matters because Amp was broken at scale before it. Now it’s not.

We’d keep paying for it if we were shipping code with a team. The sub-agents actually catch things humans miss. The AGENT.md policies enforce consistency without reinventing the wheel every PR. The token pricing is honest, even if it’s expensive.

Solo? Save your money and use Claude Code. In a team of 5+? Amp’s worth the burn rate.