Anthropic adds scheduled deployments and vaulted env vars to Claude Managed Agents

• Anthropic put claudeai's launch thread behind two new Managed Agents primitives: scheduled deployments in public beta, plus vaulted environment variables for tools that need secrets.
• The new deployment flow is effectively cron for agents. Claude's scheduled deployment docs say each job carries a cron expression, timezone, agent config, environment config, and an initial user.message that kicks off the run.
• Vaults keep API keys out of the model path. According to claudeai's vaults post, the sandbox only sees a placeholder, while Claude's vault docs say the real secret is swapped in at egress.
• Dynamic workflows are now generally available in Claude Code, where claudeai's dynamic workflows announcement says Claude writes orchestration code, fans out subagents in parallel, and verifies results before handing them back.
• The same week, ClaudeCodeLog's 2.1.172 changelog and ClaudeCodeLog's 2.1.173 follow-up showed the CLI getting deeper nesting for subagents, auto-compaction for stuck 1M-context sessions, and small cleanup fixes around model naming and Windows sandbox warnings.

You can read Anthropic's broader Managed Agents architecture post, skim the original dynamic workflows launch post, and check the exact API mechanics in the scheduled deployments docs and vaults docs. The interesting bit is how cleanly the pieces line up: agent config, environment, session log, cron trigger, and secret injection now all live in the same surface.

Anthropic's new deployment object bundles the session inputs you already need for a managed run, then adds a schedule on top.

Claude's scheduled deployment docs make the mechanics concrete:

• required: agent configuration and environment configuration
• optional: files, GitHub, memory stores, and vaults
• required kickoff: an initial user.message
• schedule shape: cron expression plus timezone
• finest interval: one minute

That makes the feature less like a generic "run later" toggle and more like a first-party recurring job runner for agent sessions.

The security model is the cleanest part of the release. claudeai's vaults post says the sandbox holds only a placeholder, not the live secret.

The corresponding vaults documentation fills in the rest:

• vaults support MCP credentials and environment_variable credentials
• each environment variable secret is keyed by secret_name
• the placeholder is substituted with the real value only when outbound traffic leaves the sandbox
• the agent never sees the underlying secret value
• environment_variable credentials are not yet supported for self-hosted environments

Anthropic has been talking about separating the harness from the execution sandbox in its Managed Agents architecture writeup. Vaulted env vars are the first feature here that makes that separation feel operational, not architectural.

Dynamic workflows moved from blog-post concept to general availability.

Anthropic's dynamic workflows post describes a specific loop:

1. Claude writes orchestration scripts during the session.
2. Those scripts launch tens to hundreds of parallel subagents.
3. Claude checks the work before anything reaches the user.
4. The feature is now GA across the Claude Code CLI, Desktop, and SDK surfaces.

The examples Anthropic gives are codebase-wide bug hunts, large migrations, and plan stress tests across legacy repos, which is a much narrower and more useful framing than "agents do bigger jobs."

The launch thread pitched bigger jobs, and the changelog immediately showed what that means in practice.

The v2.1.172 release notes added three details that were easy to miss in the marketing copy:

• subagents can now spawn nested subagents up to five levels deep
• 1M-context sessions that run without usage credits now auto-compact back to standard limits instead of stalling permanently
• /plugin got a search bar, which matters because the plugin browser was already large enough to need cursor and navigation fixes in the same release

Then v2.1.173 landed two cleanup patches a few hours later: stripping the [1m] suffix from Fable 5 model names so default long-context matching stops breaking, and suppressing a false Windows warning about missing sandbox dependencies.