Anthropic Japan BDR Automates Sales Workflow with Fleet of Claude Agents

A business development representative on Anthropic's Japan team automated his daily workflow by building a fleet of five Claude agents, collectively named Nakama ("the crew"). Previously spending hours manually routing leads and updating the CRM, he now reviews drafts prepared by agents that operate on a schedule. The transition from manual tasks to agent-assisted workflows allows him to focus on customer conversations and strategy.

The Nakama Agent Fleet

The workflow is divided into specialized agents, each with a single responsibility and a Japanese name:

• Furiwake ("sorting"): Reads the daily inbound queue, routes leads to the appropriate team member based on priority and availability, and posts a report to Slack.
• Sabaki ("handling"): Triages assigned leads, qualifies them, updates Salesforce records, and drafts first-touch emails using a shared knowledge base.
• Henji ("reply"): Scans the inbox hourly for customer emails and prepares draft replies in Japanese or English.
• Shoudan ("opportunity"): Updates Salesforce opportunity records by reading recent customer interactions.
• Nokori ("what remains"): Summarizes end-of-day meetings and drafts follow-ups based on calendar transcripts.

A sixth agent, the coordinator, manages instructions and the shared knowledge base, picking up change requests from a dedicated Slack channel every 15 minutes.

Migration to Claude Managed Agents

The representative initially prototyped the fleet using local skills and scheduled tasks in Claude. He later migrated to Claude Managed Agents (beta) to centralize management and improve reliability. Key benefits of this infrastructure include:

• Background Execution: Agents run on their deployed schedules regardless of the user's laptop status, eliminating the need for local cron jobs.
• Shared Memory: All agents read and write to a single memory store, allowing corrections made by one agent (e.g., phrasing preferences) to be learned by others.
• Secure Credentials: Access to email and Salesforce is managed via managed vaults, ensuring agents never see raw credentials and access can be revoked easily.

Handling Exceptions with Claude Tag

For tasks requiring human judgment or complex follow-ups, agents hand off items to Claude Tag (beta), a Slack-integrated version of Claude. When an agent encounters an ambiguous request, it tags @Claude in the Slack channel. Claude Tag then analyzes the thread, checks company rules, drafts responses, and sets reminders for follow-ups, only escalating to the human representative when sign-off is required.

Efficiency and Model Selection

The fleet saves approximately 60 hours of manual work per week. To optimize cost and performance, the representative uses a mixed-model strategy:

• Sonnet: Used for rule-based tasks like Furiwake and Shoudan.
• Opus: Used for judgment-heavy tasks like Sabaki, Nokori, and the coordinator.
• Advisor Strategy: Henji runs on Sonnet for speed but escalates complex decisions to Opus for advice, balancing cost with the need for high-quality replies.

The representative emphasizes that defining clear roles, maintaining a shared knowledge base, and trusting agents with routine tasks are critical for successful automation.