Amp vs Claude Code: A Two-Week Field Log (2026)

Quick answer

Amp vs Claude Code in 2026 comes down to default posture, not code quality. Both ride frontier models, so the code they write lands in the same neighborhood. Amp picks the model for you, runs tools without asking, and is built around shared team threads. Claude Code hands you the model choice, asks before it touches a file or runs a command, and is tuned for a solo operator with plan mode and rewindable checkpoints. Reach for Amp when you want a fast, opinionated agent a team can watch together. Reach for Claude Code when you want a cautious cockpit you drive alone.

I ran both on the same job for two weeks. Same repo, same branch habits, same coffee. One task I actually needed done: add Postgres full-text search to a small notes app I keep alive for myself.

The job had three moving parts. A migration to add a tsvector column and a GIN index. A search endpoint that ranks results. A debounced search box on the front end. Nothing exotic. But the migration is the scary part, and that is where the two agents stopped feeling the same.

The test, and why it was fair

Same prompt to both: "Add full-text search across note titles and bodies, ranked, with a debounced search input." Same starting commit. I read every diff. I graded on what it felt like to hand real work to each one, not on a benchmark.

Short version: both shipped working search. The generated SQL and the React were about equal. What differed was everything around the code.

Amp: opinionated, and quietly terrifying

Amp is the agent from Sourcegraph. You install a CLI and run amp, or fire a one-shot with amp -x "your prompt". The first surprise: there is no model picker. Amp exposes four modes, low, medium, high, and ultra, and it decides which model serves each one. I never chose a model the entire two weeks. Refreshing, honestly.

The second surprise is the one that matters. Per Amp's own manual, "By default, Amp does not ask for approval before running tools." So when I said "add full-text search," it wrote the migration and ran it. Against my dev database. Before I finished reading the diff.

That is thrilling for about four seconds. Then your stomach drops. It worked, the index built, the search came back ranked. But I had no seatbelt, and on a real database that default is a decision you make on purpose, not by accident.

Two Amp features earned their keep. The Oracle is a separate "second opinion" reasoning model for the harder thinking; I had it review the query plan and it caught that I was ranking title and body with the same weight. And subagents each get their own context window, so I let one draft the migration while another roughed in the UI. The shared threads and pooled team billing are clearly the pitch, but solo, that part was wasted on me.

Claude Code: cautious, and yours to configure

Claude Code runs locally in your terminal and, per its own FAQ, "asks for permission before making changes to your files or running commands." That is the exact opposite of Amp's default, and after Amp's cowboy migration it felt like putting the seatbelt back on. You also pick the model here, so I ran the heavier one for the migration reasoning and dropped down for the boring UI wiring.

Plan mode goes first. Claude Code proposed the migration, the ranked query, and the debounce, then waited. I read the plan, nudged it toward a GIN index and separate weights for title and body, and only then let it run. When one migration attempt wedged on a bad column type, a checkpoint rewind put me back one step instead of into a git archaeology session.

It is slower to first action. That is the point. On the migration, slower was correct.

The real difference is not the code

Both agents shipped full-text search that works. The output converged. The experience did not, and it split cleanly along one line: what the tool does before it does anything.

Amp's default is act. Claude Code's default is ask. Everything else follows from that. Amp routes the model so you stop thinking about it; Claude Code lets you pick so you can. Amp assumes a team watching a shared thread; Claude Code assumes one person holding the wheel. Neither is "better." They are aimed at different postures.

The trap is using the wrong one for the moment. Amp's no-prompt speed is a gift on a scratch branch and a liability on a migration against anything you care about. Claude Code's permission wall is friction on throwaway work and exactly what you want when the next command is irreversible.

How I route between them now

For my own indie work, I kept Claude Code as the daily driver, because most of my scary commands are irreversible and I like the plan-then-run rhythm. I kept Amp for greenfield spikes and the days I want an agent to just run, and for the Oracle when I need a second brain on a design call.

What tipped it for me

Honestly, the permission default. On a solo project the person who catches the bad migration is me, and I want a beat to catch it. If I were on a team living in shared threads, I think Amp's answer would win, and I would not fight it. Test both on one real task you actually care about. The demo never shows you the migration moment, and the migration moment is the whole decision.

If you want the wider field I keep reaching into, here is the roundup of AI coding agents I actually use, and if the permission question is what you are really chewing on, here is how I set up Claude Code permissions. The Hacker News thread comparing Amp with other agents is worth a read for the range of opinions.

FAQ

Is Amp or Claude Code better in 2026?
Neither wins outright. Amp is faster and more autonomous by default and is built for teams sharing threads. Claude Code asks permission before acting, lets you choose the model, and suits a solo developer who wants plan mode and rewindable checkpoints. Match the tool to the posture the work needs.

Does Amp let you choose the model?
No. As of 2026 Amp has four modes, low, medium, high, and ultra, and it routes the underlying model for each mode itself. You pick a capability tier, not a specific model. Claude Code, by contrast, lets you pick the model directly.

Does Claude Code ask before running commands?
Yes. Per Claude Code's own FAQ in 2026, it "asks for permission before making changes to your files or running commands" by default. Amp's default is the reverse: it does not ask for approval before running tools.

What is Amp's Oracle?
The Oracle is a separate, more powerful reasoning model Amp can call for a "second opinion" on complex analysis, code review, or debugging. It is slower and pricier than the main agent, so you invoke it for hard thinking rather than routine edits.

Can Amp really run without asking for approval?
Yes. Amp's documented default in 2026 is to run tools, including file edits and terminal commands, without asking. That is great for speed on throwaway work and risky on anything irreversible, so treat the first run against a real database or repo with care.

Which is better for teams versus solo devs?
Amp leans team: workspaces give shared threads and pooled billing, and multiplayer threads let people work in the same session. Claude Code leans solo: it runs locally in your terminal with permission prompts, model choice, and checkpoints tuned for one person driving.

P.S. I still flinch every time an agent runs a migration I have not read. I have decided that flinch is a feature.