In early April 2026, a code leak confirmed that Anthropic is internally testing Conway — an always-on AI agent that autonomously responds to webhook triggers without waiting for user prompts. Conway represents a fundamentally different paradigm from conventional conversational AI, shifting from reactive to event-driven execution.
Background: The Agentic AI Race Accelerates
The AI industry's competitive axis is shifting from model performance to autonomous execution agents. Microsoft launched Copilot Cowork on March 30, introducing agents that automate multi-step workflows across the Microsoft 365 ecosystem. Salesforce is deploying autonomous customer-facing agents through Agentforce in its CRM. OpenAI has published a roadmap targeting an "autonomous AI research intern" by September 2026.
Anthropic itself launched Managed Agents on April 8, reinforcing its agent infrastructure. CEO Dario Amodei has stated that 70-90% of code at Anthropic is written by Claude, underscoring how deeply agent capabilities are already embedded in the company's own operations.
What Makes Conway Different
Conway's core differentiator is its always-on architecture. While existing AI agents follow a reactive model — executing only when prompted — Conway adopts an event-driven design that automatically responds to external events such as incoming emails, database updates, or API calls.
According to leaked code, Conway features an independent sidebar UI with three panels: Search, Chat, and System. Webhooks are the architectural backbone, providing public URLs that external services can call to activate a Conway instance. The agent can natively run Claude Code, automate Chrome, execute multi-step workflows, and push notifications to users.
Anthropic is also preparing an extension ecosystem through CNW ZIP, a standard that lets developers build custom tools, UI tabs, and context handlers as Conway plugins. Given the shared infrastructure with Managed Agents, a Q2-Q3 2026 public launch appears likely.
Implications for the Korean Market
The always-on agent paradigm Conway introduces has direct implications for Voice AI. Most Voice AI today operates reactively — handling inbound calls. As always-on agents become mainstream, scenarios like automated outbound campaigns, CRM-event-triggered follow-up calls, and preemptive outreach based on real-time churn detection become practical realities.
For Korean businesses, the intersection with the AI Basic Act is critical. Always-on agents that execute autonomously without human oversight face stricter governance requirements if classified as high-impact AI. As autonomy increases, so does the need for robust governance design.
What Comes Next
Anthropic has not officially commented on Conway. However, with Managed Agents now live, Claude Code advancing rapidly, and the CNW ZIP extension ecosystem taking shape, a second-half 2026 launch is increasingly probable. The agent platform war — alongside Microsoft, Salesforce, and OpenAI — is now fully underway.