Anthropic’s Shihipar Sees Artifacts as a Way to Guide Multiple Coding Agents

A coding agent can ask what you want. Anthropic’s Thariq Shihipar thinks the harder task is helping you work out the answer while the agent is working. In a new Latent.Space interview, he described a future in which Claude Code artifacts become live interfaces for reviewing plans, commenting on work and seeing what multiple agents are doing. He cautioned that the supporting infrastructure is still being developed.

That is a proposed direction for an existing feature, not a newly launched workflow. Artifacts already turn a Claude Code session into a shareable web page. Shihipar’s interview goes further: he imagines people using such a page to guide ongoing work rather than simply inspect its progress.

Why a question is not always enough

Shihipar said Claude Code faces a difficult choice when a user gives it a task. Some people have supplied enough detail and want the agent to get on with the job; others need help clarifying what they mean. Its Ask User Question tool addresses that split, but Shihipar argued that a few answers cannot capture every preference or design detail that emerges during a project.

His example is a project dashboard, perhaps arranged like a board of tasks. Shihipar said an artifact can store persistent data in an associated database, and multiple Claude agents can access that data through an artifact connection called MCP. The dashboard could show work as it develops and provide a place for people to respond. The data connection is a capability he described; the richer dashboard workflow is one he is encouraging people to explore.

From progress page to working surface

Anthropic introduced Claude Code artifacts in June as live visual pages for work such as code walkthroughs, dashboards and checklists. They draw on the session’s codebase, connected tools and conversation. A page can be shared with teammates, updated at the same link and restored from version history. At launch, the beta was available to Claude Team and Enterprise organizations through the Claude Code command-line and desktop apps.

Shihipar’s longer-term idea gives that page another job. Instead of deciding whether to put feedback in a chat message or merely read an artifact, a user might comment on a live plan in the interface built for the current task. The same surface might show different agents’ contributions. He presented this as a direction Anthropic is working toward, not as the way every artifact functions today.

Where the agent works

The interface question connects to a second distinction Shihipar made: where the agent runs is not necessarily where it does its work. Today, he said, a user can start Claude Code locally, use remote control for some cloud behavior or run Claude Code in the cloud. He described moving toward a cloud-hosted agent that receives the user’s message and can start sessions in either environment.

He said Anthropic intends to add local computer access for such an agent over time, when the computer is online. In his description, an artifact would display work carried out by agents operating in different places; it would not itself be the agent doing that work. That distinction matters for anyone expecting the visual interface, the cloud-hosted agent and local access to arrive as one finished product.