AIToday
Large Language ModelsTop Companies' AI MovesAI Business & IndustryTop Companies AIPublished: Oct 6, 2026, 06:31 JST

Mizuho builds AI agent platform "Mizuho Cowork" with Claude Agent SDK, starts PoC

Mizuho builds AI agent platform "Mizuho Cowork" with Claude Agent SDK, starts PoC

3 Key Points

  1. What happened

    Mizuho Financial Group developed "Mizuho Cowork" using Anthropic's Claude Agent SDK on AWS, via AWS Professional Services support, and started PoC trials covering specialized tasks such as case document analysis, audit checks, and overseas report summarization and translation.

  2. Why it matters

    The platform builds on Mizuho's existing "Mizuho Wiz Base" common AI agent foundation and "agent factory" by adding a general-purpose agent that accepts "Skills" describing workflows and expertise, so it appears Mizuho can scale AI agent coverage across a wide range of operations more quickly.

  3. What to watch

    Mizuho is also considering having "Mizuho Cowork" operate browser screens for legacy internal systems without APIs, which would let the agent handle cross-system workflows without major system renovation. Whether the PoC confirms feasibility for tasks requiring specialized judgment will determine how broadly Mizuho expands the scope of adoptable operations.

WHO IT HITSInternal AI and digital talent teams at financial institutions, along with risk and compliance staff who must ensure agents operate within guardrails, are directly affected. Mizuho's approach of putting execution environments on managed AWS rather than employee PCs may also draw attention from IT architects at other regulated firms evaluating autonomous agent deployments.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

Mizuho's move builds on groundwork it had already laid: the company had established a company-wide common AI agent foundation called "Mizuho Wiz Base" and an "agent factory" to support rapid development of agents for individual operations. "Mizuho Cowork" is positioned as a further evolution of those efforts, separating Mizuho-specific business know-how from the underlying agent technology so that the latest AI models and agent technology can be continuously adopted while internal rules and expertise accumulate independently.

The architecture reflects a practical constraint Mizuho highlights in the body: autonomous agents typically run locally on user PCs, which creates challenges for enterprises needing to manage access scope and operation status. By hosting the execution environment and workspace on AWS, Mizuho avoids direct access to employees' business terminals and local data, and because agent operations do not depend on the employee's device, long-running processes can continue while the employee is not operating the terminal. The body also notes that the two-layer harness structure combines Claude Agent SDK's general agent layer with Mizuho's own layer for internal guardrails and Skills.

Looking ahead, the body suggests enterprise use of AI agents is shifting from humans instructing each step toward cloud-based agent work environments where AI autonomously assembles necessary procedures and advances long-running operations even when employees are away from their terminals. Whether Mizuho Cowork can deliver that vision is likely to hinge on the PoC results for tasks requiring specialized judgment and on how well browser operation of legacy systems works in practice, areas the body frames as ongoing verification and development rather than settled outcomes.

FAQ
What is Skills in Mizuho Cowork?
Skills is a mechanism developed and openly standardized by Anthropic to add specialized knowledge and workflows to AI agents. Mizuho uses Skills to embed each department's procedures, judgment criteria, expertise, and internal rules so Mizuho Cowork can apply them as needed.
How does Mizuho Cowork differ from conventional interactive generative AI?
Unlike conversational generative AI, autonomous agents like Mizuho Cowork operate files and tools in their work environment and execute multiple processes consecutively. Mizuho built the agent's execution environment and workspace on AWS under its own control, avoiding direct access to employees' devices or local data.
What security measures are in place?
Mizuho used Amazon ECS on AWS Fargate and Amazon S3 Files to separate execution environments and workspaces per user, and incorporated input-output controls, logging, and operational monitoring via Amazon Bedrock Guardrails. This ensures information separation while balancing scalability, cost efficiency, and the controls required of financial institutions.
Top Companies AIRead Original Article

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articlePepsiCo taps Siemens and NVIDIA digital twins, sees 20% throughput lift