New from Code with Claude Tokyo: scheduled deployments and environment variables…

Previously, Claude Managed Agents supported continuous operation but required user commands or external triggers to start. The newly announced scheduled deployment feature in public beta allows developers to configure cron expressions for agents (e.g., 'update reports daily at 10 AM'), enabling agents to automatically start and execute predefined workflows at specified times.

This means managed agents can now handle enterprise-grade tasks such as monitoring, data backup, periodic reporting, without human oversight. An official Anthropic tweet emphasized 'Agents now run on a schedule,' directly competing with traditional scheduling tools (e.g., cron jobs or AWS Lambda scheduled triggers) while encapsulating them within the AI agent's reasoning layer.

Currently, the feature is only in public beta, and it remains unclear whether it supports advanced options like automatic time zone adaptation or exception retry strategies. Nevertheless, it reveals Anthropic's ambition to embed agents into corporate infrastructure.

The second public beta feature – Environment Variables in Vaults – addresses a core security conflict when agents use third-party tools: agents need access to sensitive credentials (e.g., Slack tokens, database passwords), but traditional practice hardcodes these in prompts or configuration files, risking leaks.

Environment variable vaults encrypt and store these values on Anthropic's server side, and agents dynamically retrieve them at runtime without exposing plaintext in the generation interface or logs. This is similar to GitHub Actions Secrets or Docker Secrets, but optimized for AI agent 'tool calling' scenarios.

Enterprise users can thus implement the principle of least privilege: for example, granting an agent read-only access to a specific database, and automatically masking credentials in agent logs. However, the public beta may limit the maximum number of stored variables and the frequency of key rotation.

Combined with scheduled deployments, a clearer scenario emerges: an agent that scans AWS resources every hour can now securely use keys stored in vaults to call AWS APIs without human intervention or credential exposure.

Claude Code's dynamic workflows are the only feature that reached GA (general availability) in this update. They allow developers to define multi-agent coordination patterns: for example, a main agent can dynamically spawn several worker sub-agents, each responsible for different subtasks like code writing, testing, documentation generation, with sub-agents exchanging intermediate results through messages.

This feature was previously previewed in early versions of Claude Code, and its official release indicates that Anthropic has accumulated sufficient stability evidence. Dynamic workflows differ from LangChain's Agent Executor or CrewAI's hierarchical coordination by employing a 'dynamic' strategy – spawning and destroying sub-agents in real-time based on task requirements, rather than following a predefined fixed graph.

This is particularly well-suited for maintaining large code repositories: after analyzing the codebase, the main agent can simultaneously dispatch sub-agents for parallel fixes, refactoring, and testing, and finally merge results. The official tweet specifically mentions 'take on bigger jobs,' implying that the feature has undergone extensive internal testing.

It is important to note that dynamic workflows are currently integrated only in Claude Code, not extended to managed agents. This reflects product line differences: Claude Code, a command-line tool for developers, requires highly flexible parallel capabilities; managed agents emphasize predictability for long-term background tasks.

Choosing to announce these updates at the Code with Claude Tokyo conference in Japan is no coincidence. Japanese enterprises' extreme pursuit of automation and efficiency, along with their stringent security compliance requirements (e.g., the 2025 Data Protection Act revision), make scheduled scheduling and credential security a key entry point for Anthropic to open the Asian enterprise market.

But more noteworthy is the strategic signal: Anthropic is transforming Claude from 'generative AI chat' to a 'programmable task executor.' These three features respectively address the dimensions of time (scheduled deployments), space (environment variables), and concurrency (dynamic workflows), nearly completing all basic capabilities needed to deploy AI agents into production environments.

Currently, OpenAI's GPT-4o Agent and Google's Gemini Agent have not made similar public moves in scheduling and credential management. Anthropic's first-mover advantage may help lock enterprises into the Claude ecosystem – especially those already deeply using tools like GitHub, Jira, and Slack.