Amp Spaces delegates tasks to agents and runs up to 16 parallel threads

When several AI agents are fixing bugs, reviewing code, and preparing a release at the same time, a new problem quickly emerges: who assigns their tasks, tracks dependencies, and gathers the results? A human usually takes on this role. They create separate chats, carry context between them, and try to remember which agent is working on what.

Sourcegraph has proposed a different interface. Amp now has Spaces, shared workspaces where one agent acts as the coordinator. The user gives it a goal, and it creates separate threads, connects agents to them, and reports progress in the main chat. For a team, this feels less like talking to a single assistant and more like managing a small group of contributors.

How it works in a simple example

Suppose a team is launching a new payment system. In the shared space, a developer writes: “Check the discount calculations, prepare the migration, and monitor errors after release.” The coordinator can create three threads: one to verify the formulas, another for the database migration, and a third to monitor logs and user reports.

If a colleague discovers a coupon-rounding bug, there is no need to manually start another chat and explain the context again. The coordinator creates a thread for the bug, assigns it to a specific participant or agent, and then returns the status to the shared feed. When necessary, a person can open the thread alongside the chat and inspect the patch, discussion, or test results.

Sourcegraph says Spaces itself was developed this way. The team left comments and ideas in a shared room, and an agent distributed them among threads. Another example involved launching service accounts: from a single space, the team rolled out the feature, monitored messages on social networks and Slack, fixed bugs, sent emails, and issued refunds to affected users.

For a recurring security review, an agent divided the codebase into 16 areas, checked each one in a separate thread, rechecked the issues it found, and saved a summary for a human reviewer. This illustrates the main idea behind Spaces: parallel work remains visible without turning the main chat into a dump of intermediate messages.

Why this matters to developers and product teams

The main development here is not another button for launching an agent. Amp moves the interaction one level higher: the human manages the goal, priorities, and constraints instead of every individual conversation. This approach is especially useful for tasks that naturally break down into independent parts, such as reviewing a large repository, preparing a release, investigating bugs, or collecting feedback.

Spaces also includes a collaboration layer. People can discuss work in the shared room, reply to messages, and join specific threads. Sourcegraph says that within the company, some discussions have even moved there from Slack. This turns the space into both a dispatch center for agents and a decision log: it shows who assigned a task, how it was divided, and what the final outcome was.

Where caution is still warranted

For now, the description is based primarily on Sourcegraph's own experience. The company provides no comparative data showing how much Spaces reduces development time, how many errors the coordinator misses, or how costs increase with a large number of parallel agents. The figure of 16 refers to one security-review example, not a stated system limit.

The coordinator also does not eliminate the need for engineering oversight. If it receives overly broad permissions, it can do more than open a thread: it may deploy a change, send an email, or issue a refund. Teams still need access controls, confirmation for risky actions, tests, and a clear operations log.

The practical takeaway is straightforward: Spaces is worth trying on tasks that already contain several independent workstreams. Start with a safe scenario, such as reviewing repository modules, and define in advance which actions the agent may perform autonomously. If reviewing the results is genuinely easier than manually managing a dozen chats, the new interface saves not coding time but the attention of the person managing the process.