Anthropics works on its always-on agent with UI extensions (3 minute read)
Anthropic is building Conway, an always-on Claude agent with a containerized environment and extension system that lets users install mini-apps with custom UIs.
Deep dive
- Conway represents a shift from conversational interface to persistent agent platform with container-based runtime
- Mobile parity signals Anthropic views this as core product surface, not a desktop experiment or power-user feature
- Extension system with custom UI tabs enables mini-applications to run alongside Claude conversations
- "Installed" and "Built-in" sidebar sections suggest app launcher model similar to browser extensions or IDE plugins
- Formalizes patterns that advanced users currently build manually with OpenClaw for agent orchestration
- Each extension can ship its own interface, creating modular ecosystem of reusable agent workflows
- Container lifecycle controls let users start, stop, and manage agent instances independently
- Permissions system allows granular control over agent capabilities, critical for persistent autonomous agents
- No release timeline announced, but cross-platform development pace indicates major platform priority for Anthropic
Decoder
- Always-on agent: An AI agent that runs persistently in the background rather than only during active chat sessions
- Containerized: Running in an isolated environment with defined resources and lifecycle management
- Connectors: Integrations that allow the agent to access external services or data sources
- Webhooks: HTTP callbacks that allow external systems to trigger agent actions
- Extensions: Modular add-ons that enhance agent capabilities, potentially with custom user interfaces
- Tool calls: When an AI agent invokes specific functions or APIs to perform actions beyond text generation
Original article
Anthropic appears to be deep into development on an always-on agent internally named Conway, a containerized Claude environment that will eventually surface directly to all users in the UI. The project has been taking shape across both web and mobile builds, with recent iOS updates now carrying a full settings interface that mirrors what has already been seen on desktop. From mobile, users should be able to install extensions, manage connectors and webhooks, pick which model powers the agent, control the container lifecycle, and fine-tune tool calls, essentially full parity with the web configuration surface.
ANTHROPIC 🚨: CONWAY WILL EVOLVE ALWAYS-ON AGENTS TO THE NEXT LEVEL!
Imagine an always-on Agent with custom UI tabs that users can share and reuse as packages. Mission control, any custom workflow that requires a UI, etc.
And all these to be powered by top models from… pic.twitter.com/oeh12G3sFx
— TestingCatalog News 🗞 (@testingcatalog) April 21, 2026
Conway opens in a separate tab where users can chat with the agent, add connectors, configure extensions, and set precise permissions over what the agent is allowed to do. The codename may still shift before launch, but the scope of the build suggests it is now among the most actively developed surfaces within the company.
Recent iOS builds have picked up the full settings interface too, meaning mobile users should eventually reach parity with web, installing extensions, managing connectors and webhooks, switching the underlying model, controlling the container lifecycle, and tuning tool calls all from the phone. That is a notable commitment for something still pre-release, because it implies Anthropic wants Conway to feel like a first-class product surface rather than a desktop-only experiment.
The more intriguing detail appears in the web sidebar, where two new sections, labeled "Installed" and "Built-in," have quietly appeared. On their own, they look unremarkable, but paired with the known direction that upgraded extensions will be able to ship custom UI tabs, the setup starts to look like a launcher for full mini-applications running alongside the main Claude conversation. Each installed extension could carry its own interface that users control independently, backed by a standard format Anthropic can promote across its ecosystem, conceptually close to how Skills already function.
The addition of extensions would turn the always-on agent into a modular runtime where reusable mini-apps plug into a persistent Claude environment, covering everything from dashboards to operational mission-control panels. Power users have already been building this pattern on top of OpenClaw, wiring up custom UIs to orchestrate agent workflows; a native, packaged version from Anthropic would formalize the idea and hand every user a path to the same capability without stitching anything together themselves.
Timing remains unclear, with no public release window, but the pace of changes landing across web and mobile points to one of Anthropic's most ambitious platform moves to date.