Playbooks · Coding tasks
Amp Spaces: one agent runs the rest for you
Written November 8, 2026 · Edited November 12, 2026
Amp's Spaces put one orchestrator agent over parallel threads, plus GitLab merge requests and shared Artifacts. What a space does — and costs.
What Spaces are
Spaces, launched November 6, 2026, change who does the organizing in Amp. A space is a conversation with a single agent whose job is to spawn and coordinate agents for you. You no longer have to create and organize threads yourself. The agent in a space does it for you. Amp describes it as a higher-level way to work with many agents: you keep a bird's-eye view and can still dive into the weeds when you need to.
To start a space, use the + in the Spaces section of the sidebar, set a title and a project, and then tell your agent what to work on. The space itself is also a team room. Amp's own team moved their discussions into spaces instead of Slack, taking video calls and talking through work in the space while the agent delegated feedback and bugs into threads.
What the orchestrator actually did
Amp's launch post gives three examples from its own team, on its own codebase. Those are the vendor's own stories, not independent tests, but they show what the orchestrator is expected to do.
For a regular security review, one space was given the goal of reviewing recent changes for security issues. Its first pass split the codebase into 16 areas, reviewed each in its own thread, checked every finding in another thread, and kept a summary artifact in the space for review. The same post describes a space built for the Spaces launch itself, where the team dropped feedback and bugs into one chat and the agent delegated them into threads, and a space that ran the launch of service accounts, monitoring bug reports, fixing incoming bugs, and sending emails and refunds to affected users.
Read those as a shape, not a promise. Sixteen parallel review threads is a coordination claim. It says nothing about how many of the findings were real, how many were false alarms, or how long the review took. A space that can spawn threads without limit is also a space that can spend without limit, and nothing in the launch post puts a number on either.
What else shipped the same week
GitLab arrived November 5. Connect a GitLab account, on gitlab.com or a self-managed instance, and orbs clone and push as you, with glab signed in, so the agent can open merge requests, comment on them, and read issues and pipelines. Amp can subscribe to comments and CI for a merge request it opens, so a reviewer's request or a failed pipeline lands in the thread that wrote the code, where Amp already has the context. You can also ask Amp to start a new orb for every new merge request or issue in a project, each with a summary ready when you open the thread.
Artifacts, from November 2, are documents that belong to a single thread and that the agent, you, and your teammates can all edit. When the agent has something for you to read, like a plan, a report, or notes from a call, it writes an artifact instead of a long reply. The artifact opens next to the thread, and when the agent comes back to it, it reads your edits first and works them in instead of writing over them. The security-review summary above was kept as an artifact in its space.
The November 6 release also added click-to-fix for websites: point at any element on your site, write instructions, and send them straight to Amp. That one is a small interaction change next to Spaces, but it is the same idea: less setup between spotting something and handing it to an agent.
Strengths
- One conversation replaces a pile of threads. You give the space a goal and it does the splitting, instead of you creating a thread per task and keeping track of them.
- Review work gets a second pass by design. In the security-review example, findings were checked in a separate thread before they reached a person, which is a better shape than a single agent marking its own homework.
- The GitLab loop closes. A reviewer comment or a failed pipeline comes back to the thread that wrote the code, with the context already there, instead of starting a fresh session.
- Artifacts give the work a shared surface. A summary, a plan, or a report lives in one editable document that the agent, you, and your teammates can all change.
Weaknesses and limits
- No pricing for Spaces was in the sources read. Neither the launch post nor the changelog pages state what a space costs, which model the orchestrator runs on, or how many threads it can spawn. Do not budget a space until Amp publishes those numbers.
- The orchestrator is a single point of misunderstanding. If it splits the work wrong, every thread below it inherits the mistake. The launch post shows no example of a space going wrong, or of a person correcting a bad split.
- Parallel threads multiply cost and review load. Sixteen review threads produce sixteen sets of findings to check. The coordination is free to describe and expensive to read.
- Spaces are new. The launch post is dated November 6, 2026, and the examples are from Amp's own team during a few weeks of internal use. There is no outside report yet on how spaces behave on a codebase the team does not know.
What to try first
Give one space one bounded goal, such as reviewing the last week of changes for security issues, and watch what it spawns before giving it a second goal. Read the summary artifact it keeps, and check two or three findings yourself against the code before trusting the rest. If the split looks wrong, say so in the space and see whether the orchestrator corrects it or keeps going. That first goal is the test: it tells you whether the coordination saves you time or just moves the organizing from your hands to a summary you still have to read.
Limits
- Nothing here reports a Spaces run. The behaviour, the examples, and the gaps above come from Amp's news posts and changelog, and no space was opened while writing this page.
- Pricing, model choice, and per-thread limits for Spaces were not in the sources read. The page states that gap rather than filling it.
- The dogfood examples are Amp's own team on their own codebase. They show the intended shape, not an independent measurement.
Related playbooks
- Bring your own agent into Android Studio, and switch when its quota runs out
- Claude Code 2.1.319: unattended runs stop stalling
- Cloudflare runs the Codex loop in a Durable Object
- Foundry Local runs Qwen3.8-27B on your GPU
Sources
Back to: Coding playbooks hub · Also: Codex 0.167: run parallel sessions in one place · T3 Code: drive your agents from your phone · Run parallel subagents in Claude Code