Amp Review 2026: Sourcegraph's Agentic Coding Agent Tested

Amp is one of the strongest agentic coding tools we've tested in 2026, and it's the one we'd reach for on messy, multi-step work that spans several files and skills. Sourcegraph's pitch is simple: don't bet on a single model. Amp routes each subtask to the best-suited frontier model, spawns parallel subagents, and escalates the hard parts to a dedicated reasoning model called the Oracle. It runs everywhere you code — CLI, web, and IDE — and it bills pay-as-you-go from a credit balance instead of a subscription. That last part is both its best feature and its biggest catch.

What is Amp?

Amp is Sourcegraph's agentic coding tool — the direction the company took after Cody. It's built to plan and execute multi-step engineering work on its own, not just answer questions or complete a line. You give it a task and it reads your code, makes a plan, edits files, runs commands, and iterates.

The core idea that sets it apart is that Amp doesn't tie you to one model. It ships as a CLI, a web UI, and IDE extensions for VS Code, JetBrains, Neovim, and Zed, and underneath it decides which frontier model should handle each piece of a job. It's a strong contender in our best AI coding assistants ranking, especially for teams.

Key features

Multi-model routing

Amp's routing layer sends each subtask to the model best suited for it, rather than forcing everything through one. You pick from four modes — Low for fast, cheap small tasks, Medium for balanced work, High for deep reasoning, and Ultra for the most capable, open-ended jobs. The context window runs to about 200K tokens, so it holds large chunks of a codebase in view. In practice this routing is why Amp edges out single-model agents on mixed workloads: it doesn't waste a top-tier model on a rename, and it doesn't hand a hard refactor to a cheap one.

Subagents & the Oracle

For anything multi-step, Amp automatically spawns parallel subagents that tackle pieces of the work at the same time, then reconcile their results. When a task needs real analysis — architecture calls, gnarly debugging, tricky design trade-offs — it escalates to the Oracle, a separate, more powerful reasoning model. Cheaper models keep the routine work moving while the Oracle handles the parts that actually need thought. That split is where Amp feels most like a senior engineer rather than an autocomplete on steroids.

Surfaces, Amp Tab & collaboration

Amp works across the CLI, web, and four IDEs, and you can start a thread in one and continue it in another. Amp Tab is its inline autocomplete, so you get keystroke-level suggestions alongside the full agent — a nice touch that agents like OpenCode skip. It supports MCP for custom tools, and it leans hard into team use: thread sharing, workspace collaboration, and even leaderboards. If you run a team, that shared-context layer is a real differentiator.

Output quality

On quality, Amp is very good, and the routing is the reason. Because the Oracle and High/Ultra modes come in for the hard parts, its output on complex, open-ended tasks was more reliable than single-model agents we've run on the same problems.

The parallel subagents also help with breadth — Amp is comfortable touching many files at once and keeping the changes coherent. It's less about raw speed on a single small edit and more about handling a whole feature end to end. The honest caveat is polish: Amp is newer than Cursor, and Cursor's IDE experience still feels more refined for tight edit-and-review loops. Amp's strength is the agent, not the surrounding UX.

Pricing & credits

Amp is pay-as-you-go with no required subscription — the headline is that there's no hard token cap and, for individuals, zero markup on provider API pricing. You top up credits (minimum $5) and Amp deducts your actual LLM usage. A thread that burns $2 of Anthropic tokens and $0.50 of OpenAI simply deducts $2.50. New users get about $10 in free credits to try it, and credits pool across a workspace, expiring after one year of inactivity.

Plan	Price	What you get
Pay-as-you-go	From $5	Credits at zero markup; actual usage deducted; no hard cap; ~$10 free to start
Megawatt	$20/mo	$20 agent usage included, 750 hours of small async "orbs," Low & Medium modes only
Gigawatt	$200/mo	$200 agent usage included, 1,000 hours of large orbs, all modes incl. High & Ultra
Enterprise	Usage-based	~50% more than individual; BYO inference keys, SSO, zero data retention

The credit model is the whole story, good and bad. It means you're never throttled mid-task and you pay exactly for what you use — but pay-as-you-go can surprise you, and there's no flat "unlimited" plan to hide behind. For a full breakdown, see our Amp pricing guide.

Pros & cons

Who it's for

Use Amp if you want an agent that picks the right model for each subtask, never hits a hard cap, and plays well with a team — the routing and subagents genuinely pay off on complex, multi-file work, and the pay-as-you-go pricing means you never buy more than you use. Skip it if you want a predictable flat bill or the most polished IDE loop: Claude Code is simpler to budget on a subscription, and Cursor has a more refined editor experience.

Weighing Amp against the obvious rival? Read Amp vs Claude Code, check the Amp pricing breakdown, or see the full best AI coding assistants ranking. You can also try it directly on Amp's site.