At Code with Claude Tokyo in June 2026, Anthropic shipped three production-oriented upgrades that turn Claude from a chat interface into scheduled, credentialed, parallel infrastructure—and the developer discourse on X framed the same week as a shift from prompt engineering to loop engineering.
@claudeai summarized the launch:
New from Code with Claude Tokyo: scheduled deployments and environment variables in vaults are in public beta in Claude Managed Agents, and dynamic workflows in Claude Code are generally available. Agents now run on a schedule, use your tools securely, and take on bigger jobs.
On June 11, @ClaudeDevs reiterated both Managed Agents features for developers building on the Claude API:
We've just added two new Claude Managed Agents features: 1. Scheduled deployments — run tasks on a schedule. 2. Environment variables — expose vault credentials for CLIs as environment variables.
Onboard from Claude Code with /claude-api managed-agents-onboard — covered in our Claude Code commands reference.
TL;DR
| Feature | Status | What it does |
|---|---|---|
| Scheduled deployments | Public beta (Managed Agents) | Cron-fired agent sessions—daily reports, compliance scans, monitoring |
| Vaults (env vars) | Public beta (Managed Agents) | API keys as opaque placeholders; injected at network egress |
| Dynamic workflows | GA (Claude Code) | Parallel subagent orchestration for large codebase jobs |
| Philosophy shift | Industry-wide | Build systems that prompt themselves; verify the right work |
From prompting to loop engineering
An Anthropic engineer quote circulated widely on X via @Rixhabh__:
You're not supposed to prompt Claude. You're supposed to build a system that prompts itself.
That matches what @ClaudeDevs said when Fable 5 launched:
We used to verify that Claude did the work right. Now we verify that it's doing the right work.
The Tokyo announcements are infrastructure for that second model:
| Old pattern | New pattern |
|---|---|
| Open Claude, ask once, leave | Scheduled deployment fires daily |
| Paste API keys into prompts | Vault holds secrets at egress |
| One agent, one thread | Dynamic workflow fans out subagents |
| Human drives every step | Human sets goal, schedule, rubric |
Thariq's Fable-edited launch video is the creative proof; Tokyo is the platform proof.
Scheduled deployments: cron for managed agents
Scheduled deployments give a Claude Managed Agent a cron schedule and timezone. When the schedule fires, the agent starts a new session and completes its task—no scheduler for you to build or host.
From Anthropic's announcement:
Each time the schedule fires, the agent starts a new session and completes its task, with no scheduler for you to build or host.
Example use cases
| Use case | Cron pattern | Initial message |
|---|---|---|
| Weekly compliance scan | 0 20 * * 5 (Fri 8pm ET) | "Run the weekly compliance scan." |
| Daily engineering report | 0 9 * * 1-5 | "Summarize yesterday's PRs and CI failures." |
| Usage monitoring | Hourly / daily | "Flag customer usage surges for review." |
Anthropic documents manual runs via a run endpoint—test before committing to cron. Org limit: 1,000 scheduled deployments per organization (contact support for more).
API: Requires beta header managed-agents-2026-04-01. See scheduled deployments docs.
Vaults: secure API keys for CLIs and MCP
Managed agents connect through API calls, CLIs, and MCP. Vaults now store environment variables so CLIs authenticate without the agent ever seeing the secret.
How it works (vault docs):
- Register a credential with a secret name (e.g.
NOTION_API_KEY) and allowed domains. - The sandbox holds an opaque placeholder, not the real key.
- On outbound HTTP requests to approved domains, the real secret attaches at the network boundary.
- Rotate keys in the vault; running sessions pick up new values on the next call.
Supported credential types:
| Type | Use case |
|---|---|
environment_variable | CLIs, SDKs, direct API calls |
mcp_oauth | MCP servers with OAuth |
static_bearer | Bearer-token MCP servers |
Compatible CLIs named by Anthropic include Browserbase, KERNEL, Notion, Ramp, and Sentry. Browserbase and KERNEL add browser capabilities to managed agents for the first time—agents can navigate the web alongside other tools.
Ramp example (from Anthropic)
Ramp stores credentials in vaults so a managed agent can monitor usage and customer conversations, flagging surges as they happen so the team can confirm whether activity is intended—ideal for pairing with scheduled deployments.
Dynamic workflows: GA in Claude Code
Dynamic workflows graduated from research preview to general availability at Tokyo. Claude Code can now generate scripts that run tens to hundreds of parallel subagents for jobs that do not fit a single thread:
- Codebase-wide migrations and framework swaps
- Bug hunts and security audits
- Profiler-guided optimization passes
- Work spanning hundreds to thousands of files
Each subagent works independently; results are checked before merge; the workflow iterates until convergence. Progress persists across interruptions—jobs pick up where they left off.
Trade-off: Substantially higher token usage than a normal session. Anthropic recommends scoping the first workflow and confirming before large runs. Enable via ultracode in the effort menu or ask Claude to create a workflow directly.
This complements /batch and /loop patterns in the CLI—Managed Agents for cloud cron, Claude Code for local parallel orchestration.
Rakuten and the expanding Claude ecosystem
@goyalshaliniuk and others on X reframed the same week: Claude is an ecosystem, not a chatbot—Cowork, Code, Managed Agents, MCP, plugins, and API.
Concrete production mention from Anthropic's materials: Rakuten uses managed agents for spreadsheet analysis—the kind of recurring, tool-heavy task scheduled deployments target.
This builds on May 2026's Dreaming, multiagent orchestration, and Outcomes betas. Tokyo adds when (schedule) and how safely (vaults) agents run in production.
Risks and human review
GroK-summarized discourse and practitioner threads flag real constraints:
| Risk | Mitigation |
|---|---|
| Token cost on dynamic workflows / long schedules | Scope cron jobs; set budgets; review first runs manually |
| Wrong work done well | Outcomes rubrics, verifiers, human review of artifacts |
| Secret scope creep | Domain-allowlist vault credentials; audit vault access |
| Scheduled drift | Manual run tests; monitor session outputs |
Loop engineering does not remove humans—it moves them to system design and verification, as in Fable loop design.
How to get started
Managed Agents (schedule + vaults):
- Read Claude Managed Agents docs
- Create an agent and environment in Claude Console
- Register vault credentials with domain allowlists
- Create a scheduled deployment with cron + initial
user.message - Use beta header
managed-agents-2026-04-01
Claude Code (dynamic workflows):
- Upgrade to a plan with workflow access (Max / Team / Enterprise per Anthropic docs)
- Enable ultracode or ask Claude to orchestrate a workflow
- Start scoped—one migration or audit before quarter-scale jobs
In Claude Code, run /claude-api managed-agents-onboard or ask to start onboarding for managed agents in Claude API.
Related explainx.ai guides
- Claude Managed Agents: Dreaming and multiagent orchestration
- Dynamic workflows in Claude Code
- Loop engineering guide
- Claude Code commands reference
- Fable 5 edited its own launch video
- DiffusionGemma 4× faster text generation
- What is MCP?
Primary sources: What's new in Claude Managed Agents · @claudeai Tokyo announcement · Scheduled deployments API · Vaults API
Summary
Code with Claude Tokyo delivered the missing ops layer for agentic AI: cron schedules for recurring work, vaults for CLI/MCP secrets, and dynamic workflows GA for parallel codebase jobs. Together with Fable 5 and loop primitives, Anthropic's story is coherent—design the loop, store the credentials, schedule the run, verify the outcome.
Bookmark this alongside our managed agents deep dive and loop engineering guide when building production agent systems.
Feature availability, beta headers, and pricing reflect Anthropic announcements as of June 11, 2026. Verify claude.com/blog and platform docs before production deployment.