Anthropic is quietly building something that looks a lot like the AI agent everyone keeps promising but nobody has shipped yet. Conway, an unreleased always-on agent built on Claude models, appears designed to operate persistently in the background, responding to real-world triggers without a human babysitting the chat window.

The project surfaced through a code leak earlier in 2026 involving over 500,000 lines of code. Anthropic hasn’t publicly acknowledged Conway’s existence, which makes the level of detail available all the more interesting.

What Conway actually does

The leaked code references a specialized “Conway instance” with its own user interface, webhook listeners for real-time events, GitHub integrations, and push notification capabilities. In English: it can watch for things happening across your tools and services, then take action automatically.

The core feature set includes always-on persistence, event triggers tied to external services like GitHub, scheduled task execution, and browser automation. Browser control means Conway could theoretically navigate websites, fill out forms, pull data, or interact with web applications the same way a human would.

This is a fundamentally different product from Claude Code, Anthropic’s existing developer-focused tool that runs sessions on a one-off basis.

The breadcrumbs Anthropic left in plain sight

Around April 16, 2026, the company introduced “routines” for Claude Code, a feature enabling event-triggered agent tasks running on managed cloud infrastructure. Routines let developers set up workflows that fire when specific conditions are met, essentially a lighter version of what Conway appears to offer as a full product.

Earlier, a research paper published on February 18, 2026, revealed that the longest-running Claude Code sessions had increased in duration from under 25 minutes to over 45 minutes within a three-month span.

No confirmed launch timeline or public testing access has been established.

Why this matters for the AI agent race

Conway positions Anthropic directly against competitors building their own persistent agent platforms. The webhook and GitHub integration angle suggests a developer-first approach, which makes sense given Claude Code’s existing footprint in engineering workflows. But browser automation opens the door to a much broader user base.

The crypto angle here is essentially nonexistent. Nothing in the leaked code or public signals connects Conway to blockchain technology, digital tokens, or decentralized infrastructure.