The months when "building your own agents" was standard practice
From late 2025 through early 2026, many Japanese companies struggled with how to build Claude-powered agents in-house. Assembling agent loops, sandboxes, tool execution layers, and session management on top of the Messages API required months of steady, specialized infrastructure work.
Then, on April 8, 2026, Anthropic released Claude Managed Agents in public beta. The ant CLI for resource definition was also released on the same day, accompanied by the phrase "infrastructure abstracted away" throughout the official release notes. The message is clear: "We will take care of the entire infrastructure needed to run agents."
To translate this announcement into actionable decisions for Japanese enterprises, this article examines the four core concepts, the pricing model, and the lineup of early adopters, concluding with an analysis of whether to continue custom builds or pivot to managed services.
1. What changes with Claude Managed Agents?
Claude Managed Agents is a fully managed harness designed to run Claude as an autonomous agent. Developers need only define the following four concepts; Anthropic's infrastructure handles session management, tool execution, and streaming.
| Concept | Role |
|---|---|
| Agent | A definition bundling model, system prompt, tools, MCP servers, and Skills. Once created, it can be reused by ID |
| Environment | Container templates pre-installed with Python, Node.js, Go, and more. Network controls and file mounting are also supported |
| Session | An execution instance launched by specifying an Agent and an Environment. One session per task |
| Events | Messages between the application and the agent. User inputs, tool results, and status updates are streamed via SSE |
Built-in tools including Bash, file operations, web search/fetch, and MCP connectivity are available out of the box, and optimizations such as prompt caching and context compaction are handled automatically by the harness. The division of roles with the Messages API is clear.
| Messages API | Managed Agents | |
|---|---|---|
| Positioning | Direct model access | Managed agent execution environment |
| Ideal applications | Short-duration tasks requiring custom control | Long-running, asynchronous autonomous tasks |
| Infrastructure management | Built in-house | Entirely handled by Anthropic |
Anthropic's proposition is straightforward: choose the Messages API for short conversational interactions, and transition to Managed Agents for long-running batch processes.
2. Pricing model — What an architecture with "no idle charges" means
Pricing is based on token consumption plus the following pay-as-you-go fees.
- Active runtime: $0.08 per session hour (calculated in milliseconds)
- Web search: $10 per 1,000 queries
- Idle time is not billed (time spent waiting for the next user input or tool response)
The final sentence succinctly illustrates the design philosophy. Because no charges accrue while an agent is waiting, this pricing structure is designed for asynchronous workflows with substantial wait times, revealing Anthropic's targeted use cases. The primary battlefield is "slow-running agents" that execute over minutes to hours.
Conversely, Managed Agents is overkill for single-request chat responses. Failing to recognize this distinction risks concluding that sticking with the Messages API would have been sufficient.
3. Use cases demonstrated by Notion, Rakuten, and Asana
The profile of the three companies highlighted as early adopters clearly reflects the intended goals of Managed Agents.
| Company | Business domain | Anticipated agent profile |
|---|---|---|
| Notion | Productivity SaaS | Agents for researching, summarizing, and rewriting across extensive document collections |
| Rakuten | E-commerce, finance, telecommunications | Customer support and merchant assistance spanning product data and customer histories |
| Asana | Project management | PM agents that interpret task dependencies to organize progress and extract risks |
What they have in common is the pattern of "embedding" agents into existing proprietary workflows. Rather than building AI products from scratch, the use case centers on connecting Claude to internal data, tools, and MCP servers to run sustained operations.
This aligns with Anthropic's concurrent announcement of Claude Mythos Preview and Project Glasswing, as well as OpenAI's moves involving Codex pricing revisions and the Industrial Policy Blueprint, making it clear that AI vendor strategies are shifting from selling models to selling agent operational infrastructure.
4. The "build vs. managed" decision Japanese enterprises must make now
Should Japanese enterprises abandon custom agent builds? While the answer is case-by-case, the evaluation criteria can be cleanly structured.
| Condition | When to continue custom builds | When to lean toward managed |
|---|---|---|
| Data residency | Domestic data center is mandatory | Cross-border is acceptable / primarily public data |
| Runtime requirements | Proprietary GPUs / dedicated environments needed | Standard Python/Node.js suffices |
| Phase | In production with established SLAs | PoC / experimental validation phase |
| Task duration | Sub-second to second-level real-time response | Asynchronous processing taking minutes to hours |
| Internal expertise | Dedicated agent infrastructure operations team present | Infrastructure talent is scarce |
Projects combining all three conditions—PoC phase, asynchronous tasks, and limited infrastructure talent—stand to gain immense value from Managed Agents. For enterprises planning to take agents built during 2025 internal hackathons into production in 2026, now is an ideal time to evaluate migration.
Meanwhile, with the emergence of multi-agent operating systems like OpenCrew, selecting infrastructure increasingly boils down to deciding what to retain control over internally. Rather than going all-in on custom or all-in on managed, the pragmatic solution lies in separating workloads by specific requirements.
5. Migration and validation checklist
For practitioners looking to experiment firsthand, here are five verification steps.
- Prepare beta headers — All Managed Agents endpoints require the
managed-agents-2026-04-01header (automatically included in official SDKs) - Define a small agent — Start with a minimal setup of one model plus two to three tools, expressed in natural language or YAML via the
antCLI - Design the Environment — Configure with least privilege by narrowing down necessary packages, allowed network scopes, and mounted files
- Monitor session billing — Measure initial sessions to verify whether
$0.08/セッション時is billed as expected - Connect to internal MCPs — Turn existing internal APIs into MCP servers and test invocation from Claude
Steps 3 and 4 are frequent blind spots during production migrations. Stuffing an Environment with extraneous packages slows down container launch, prolonging active runtime and inflating billing. Keeping "least privilege, minimal runtime" in mind can easily change monthly costs by several multiples.
Conclusion
The announcement of Claude Managed Agents is not merely a new feature release; it signals that Anthropic's core business focus has shifted from "model provider" to "agent operations infrastructure provider." As OpenAI pushes ahead with its industrial policy blueprint and Google advances Gemini API tiering, Anthropic has staked out the next battleground by hosting the entire execution environment for agents.
The question for Japanese enterprises is simple: build on top of this platform, or continue maintaining your own infrastructure? Although the answer varies across projects, organizations can no longer afford to delay the question itself. Testing at least one proof-of-concept pipeline represents the minimal investment needed to avoid regrets in late-2026 decision-making.
References
- Claude Managed Agents overview — Claude Platform Docs
- Claude Platform Release Notes (April 8, 2026)
- With Claude Managed Agents, Anthropic wants to run your AI agents for you — The New Stack
- Anthropic Unveils Managed Agents for Claude, Eyeing Enterprise AI Workflows — Startup Fortune
- Anthropic Launches Claude Managed Agents Platform for Enterprise AI Deployment — Blockchain.news
