Claude agent fleet for sales: how a Japan BDR runs his day

An Anthropic Japan BDR runs five Claude Managed Agents plus a coordinator for leads, email and CRM updates, and counts about 60 hours saved. What to copy.

A business development representative (BDR) on Anthropic's Japan team runs five scheduled Claude agents and a coordinator. They route inbound leads, draft customer emails in Japanese or English and keep Salesforce current. By his own weekly log, "the fleet saves me around 60 hours of work". Anthropic published the account as How an Anthropic Japan BDR runs his day with a fleet of Claude agents.

This site is independent and not affiliated with Anthropic. We read the article on 2026-10-16 and did not build the fleet. The one thing we ran ourselves is the reply-drafting prompt further down.

The crew: one job per agent

He calls the group the Nakama, "the crew". Each agent has one clear job and a Japanese name, because they mostly work in Japanese.

Furiwake ("sorting") reads the daily inbound queue, routes each lead to a rep and posts a report to a Slack channel.
Sabaki ("handling") qualifies the leads assigned to him, updates Salesforce and drafts the first-touch email.
Henji ("reply") scans his inbox hourly and drafts replies in Japanese or English, depending on the thread.
Shoudan ("opportunity") pushes recent customer interactions into Salesforce.
Nokori ("what remains") runs at the end of the day. It reads the transcripts of the meetings on his calendar, summarizes them and drafts the follow-ups.
The coordinator edits the other agents' instructions when he asks for a change in plain English.

His design rule: "I divided up my workflow across several agents and gave every agent one clear job." A single responsibility keeps each agent easy to reason about when something goes wrong.

Start in Claude, move to Managed Agents later

He did not begin with Managed Agents. He prototyped the whole fleet in Claude first, with three pieces:

A skill per agent: the instructions for one job.
A scheduled task per skill, so nothing runs by hand.
A shared folder on his laptop as the knowledge base. Every skill writes back the difference between what it drafted and what he sent.

On Managed Agents, the skill became the agent's system prompt, the scheduled task became a schedule on the agent's deployment, and the shared folder became a memory store. His advice: "I'd suggest starting in Claude and moving your agents to Managed Agents once you have more than you can comfortably keep track of."

If skills are new to you, our Claude Skills guide covers what one is and how to write the first.

What Managed Agents added

He names three things that mattered most:

Background runs. "The agents run on their schedule whether my laptop is open or not, even when it's offline." He sets the cadence and timezone in the Console, so no cron jobs to maintain on the side.
One shared memory. Agents compare their drafts with the emails he sent and save each difference as a note. "I fix a phrasing once in Henji's drafts, and Sabaki and Nokori stop making the same mistake."
Scoped credentials. Credentials live in managed vaults. "Each agent's deployment uses only the vaults its job needs, it never sees the credentials, and I can revoke access in one step."

The vaults page says "The agent never sees the secret value." Our secure Managed Agents guide lists the settings to check before an agent touches customer mail.

Claude Tag: the hand-off to Slack

Some items need more than a draft: an upgrade path, a purchasing route, an issue for another team. The agents hand those to Claude Tag (in beta), which is Claude set up as a member of the team's Slack workspace.

An agent posts the action item in the Nakama channel and tags Claude. Claude reads the thread, checks the rules and terms that apply, lays out the options and drafts what is needed. When a customer or another team must reply first, it sets itself a reminder and checks back on the day.

The Claude moves worth stealing, by email

Workflows tested on real jobs, sent only when they are worth it. Free.

Which model runs which agent

He does not run everything on the biggest model. Furiwake and Shoudan follow a written rulebook, so they run on Sonnet. Sabaki, Nokori and the coordinator run on Opus because their work needs more judgment. The coordinator needs it most: it edits the other agents' instructions, and one bad edit would reach every agent.

Henji replies all day and every customer's email differs. Opus every time would cost the most, so Henji runs on Sonnet and asks Opus for advice when it hits a decision it cannot make. For how the two models differ, see Claude Opus vs Sonnet.

Try one agent yourself this week

You can copy the pattern with one inbox and no Managed Agents. This is our version of Henji, built from what the article describes: drafts only, a summary on top, language by thread, and pricing questions, sensitive accounts and ambiguous asks routed to a person. Replace the two sample threads with your own.

You are my reply-drafting assistant. Draft only, never send.
Rules:
1. Start with a two-line summary of the thread, in English.
2. Reply in the language of the thread (Japanese or English).
3. For pricing questions, sensitive accounts or ambiguous asks, write
   "Needs a human:" and one sentence on why. Do not draft a reply.
4. Never invent availability, prices or commitments. Put anything I must
   fill in inside [brackets].

Thread 1:
山本: 先日はありがとうございました。来週の火曜日か水曜日の午後に、30分ほどオンラインでお時間をいただけますか？

Thread 2:
Priya: Could you send your best price for 60 seats, and tell us whether annual billing gets a discount?

Output excerpt from our run (both threads are invented samples):

Thread 1
Summary: Yamamoto thanks you for the earlier conversation.
He asks for a 30-minute online call next Tuesday or Wednesday afternoon.
Draft:
山本様

ご連絡ありがとうございます。
来週の火曜日または水曜日の午後で調整いたします。[空き状況を確認のうえ、候補日時を2、3つ]お送りしますので、しばらくお待ちください。

どうぞよろしくお願いいたします。
[署名]

Thread 2
Summary: Priya asks for a best price for 60 seats.
She also asks whether annual billing gets a discount.
Needs a human: this is a pricing question.

Tested on claude-sonnet-5-5, October 2026. The test shows that the prompt follows its four rules on two short threads. It says nothing about running a fleet. The first thread routed to a draft with a placeholder instead of an invented time, and the second routed to a person.

What to copy, in order

now

Pick one routine job

Something you repeat daily, such as first replies to inbound mail.

Write its lane

What the agent may draft, and what always goes to you.

Run it by hand and fix the draft

Describe the job once, correct the draft, save the correction.

Schedule it only when fixes get rare

A scheduled agent runs whether or not you are watching.

Add a second agent

Share the notes between them instead of lengthening prompts.

This sequence is our reading of his lessons: "You don't need a detailed prompt every time" and "The knowledge base matters more than the prompt." It is not a procedure he published.

Where this falls short

The number is self-reported. The 60 hours comes from his weekly log, "counting each task as the time it would have taken me to do manually". One person counted it, and nobody measured it independently.
It is a vendor's own case. An Anthropic employee wrote about Anthropic's products, and Anthropic published it. Treat it as a worked example, not a benchmark.
He still reviews. CRM records stay current "(mostly)" without his input, and his role is "to review anything that requires human judgment".
Customer mail is untrusted input. Our secure-deployment guide gives the rule for agents that read it: "The rule is not to give one agent both." Read it before you give an email agent write access.
Beta pieces. Claude Tag is marked beta, so details can change.

Dynamic workflows, the Managed Agents feature that runs many agents in phases, is a different tool from this fleet. See our dynamic workflows guide. If your pain is a plain Outlook inbox, start with Claude for Outlook first.

Want the next write-up like this one? Our newsletter sends them.

Do I need Claude Managed Agents to build an agent fleet?

No. The BDR prototyped the whole fleet in Claude first, with one skill and one scheduled task per agent and a shared folder as the knowledge base. He suggests moving to Managed Agents once you have more agents than you can comfortably keep track of.

How many hours does the fleet really save?

About 60 hours, by his own count. He added up a weekly log of what the agents handle, counting each task as the time it would have taken him manually. It is a self-reported figure from one person, not a benchmark.

Does the fleet send emails without a human?

The drafts wait for him. Each email draft arrives with a short summary on top, and he reads it, confirms, revises and sends. Pricing questions, sensitive accounts and ambiguous asks route to him.