Artifacts in Amp Consolidate Findings From 60 Threads Into One Editable Document

A long conversation with an AI agent quickly turns into a storage room: the plan is somewhere above, the research findings are farther down, and the corrected table appears another twenty messages later. Making a decision requires reconstructing the context. Sourcegraph has added Artifacts to Amp: standalone documents within a thread that the agent, the user, and the user's colleagues can edit together.

This is more than an enlarged message. An Artifact opens alongside the conversation and functions as a shared working document. Users can rewrite a paragraph, mark an item as complete, modify a table, or clarify a requirement. When the agent returns to the task, it first reads the edits made by people and continues working with those changes in mind.

What It Looks Like in a Real Task

The Amp team demonstrated an investigation into the cost of model calls while a long-running terminal command was executing. The user wanted to determine whether the agent was consuming resources while merely waiting for the process to finish. The agent reviewed 60 threads, estimated the cost of the current waiting approach, explained the causes, and proposed a two-stage remediation plan.

The result appeared not as a long series of replies but as a single Artifact. The document contained findings, a diagram, and an action plan. According to Amp's illustration, the status polling currently in use required about 3,900 model calls, compared with 978 for an asynchronous monitoring approach. This is not a promise of fourfold savings for every project, but a figure from one specific internal study. The format itself is useful, however: an engineer can verify the calculations, a product manager can discuss priorities, and a manager can understand the issue before choosing a solution.

The second example concerns prototype comparison. For a Markdown preview, the agent built two alternatives in parallel in separate workstreams. It then placed them side by side in one Artifact, showing screenshots, affected files, test results, implementation cost, and a recommendation on which option to release first. Instead of a retelling of two branches, the team received an editable decision memo.

What Changes in Team Workflows

The main benefit of Artifacts is the separation of the conversation from its result. Chat remains the place for questions, commands, and clarifications, while the document becomes the current version of a plan or report. This is particularly convenient for tasks in which an agent investigates dozens of files, runs several experiments, or aggregates results from parallel threads.

For developers, such a document preserves verifiable conclusions: which files will change, which checks have already been completed, and which risks remain. A product manager can change a priority or readiness criterion directly in the table. Another participant does not need to read the entire history of the interaction with the agent; opening the final material is enough, with the original thread available when needed.

An Artifact can be created through a direct request. Amp can also select this format on its own when preparing a plan, report, or meeting notes. All documents in a thread are available under Views → Artifacts.

Limitations and Practical Takeaway

Artifacts do not eliminate model errors or turn an agent's output into verified fact. The shared document still needs to be reviewed: numbers, sources, test results, and assumptions must be checked. If a team edits one document without restraint, it also needs clear rules about who makes the final decision and which items count as approved.

The practical workflow is simple: ask the agent to put the implementation plan, option comparison, and research summary into an Artifact. Record readiness criteria, links to evidence, cost, and open questions in the document itself. After human edits, return the task to the agent and have it continue from the updated version. This makes the Artifact not a decorative add-on to chat, but a shared assembly point for the decision.