Run a Claude Trading Agent on a Schedule (Managed Agents, Cron + Vault)

What You Are Building

A Claude agent that runs on a fixed schedule in Anthropic’s cloud, not on your laptop. On June 9, 2026 Anthropic put two Managed Agents features into public beta: cron-style scheduling and a credential vault. Together they let you set an agent to fire every morning, pull market data through a connector, and hand you a research brief before the open.

What Changed in June 2026

Until now, an automated Claude trading workflow meant keeping a machine awake and a crontab entry pointed at the Claude Code CLI. That works, but it ties the job to one always-on box and leaves your API keys sitting in a .env file next to the script.

The Managed Agents beta moves both pieces into Anthropic’s platform:

Each scheduled run starts a new session and executes the task from scratch, so there is no drifting state to debug. The vault stores secrets so the agent reaches a tool without the key ever entering the model’s context. Both features are in public beta for Claude for Work and enterprise plans.

If you have wired up the local version before, the daily AI trading research routine covers the manual workflow this replaces.

What a Scheduled Agent Is Good For (and Not)

Be honest about scope before you build. A scheduled agent runs once per trigger and ends. That fits jobs that produce a report or a single decision, not jobs that need to watch a price tick by tick.

Good fits:

• A pre-market research brief: overnight news, rating changes, and filings on your watchlist.
• A daily or weekly portfolio summary mailed to you.
• A scheduled DCA or rebalance check that places one order against your rules.
• A weekly scan of a strategy’s logs for anything that broke.

Poor fits:

• Anything needing sub-minute reaction to price. A cron tick every few minutes is not a market data feed.
• Stateful strategies that must remember intraday context between runs. Each run starts clean.

For the watch-the-tape kind of bot, you still want a long-running process like the Hyperliquid trading bot or a grid bot.

Prerequisites

• A Claude for Work or enterprise plan with Managed Agents beta access
• A broker or data connector you can authenticate (Alpaca, Bybit, or an MCP server)
• API keys for that connector
• A task you can describe as a single self-contained job

Step 1: Write the Agent Task

A Managed Agent runs the same way each time, so the task description has to be self-contained. It cannot rely on anything you typed in an earlier session. Write it as a standing instruction.

Here is a pre-market brief task. Save it as task.md in your agent’s directory:

# Daily Pre-Market Brief

You run once each weekday at 7:00 AM ET. Do the following and stop.

1. For each ticker in watchlist.txt, pull overnight news and any
   analyst rating changes from the connected data source.
2. Check for new 8-K or earnings filings since yesterday's close.
3. Summarize each name in three lines: what changed, why it matters,
   and whether it shifts the existing thesis.
4. Flag anything that gapped more than 3% pre-market.
5. Email the brief to the address in config.

Do not place any trades. This is a research-only job.

The last line matters. Keep research agents and execution agents separate so a scheduling mistake can never fire an order you did not intend.

Step 2: Store Credentials in the Vault

Instead of putting your data-source key in a file, register it in the vault. The agent references the secret by name, and the platform injects the value when the agent calls the connector. The model never sees the raw key.

In the Managed Agents console, add each secret:

In the task or connector config, reference them by name rather than value:

connectors:
  market_data:
    type: alpaca
    key_id: ${vault:ALPACA_KEY}
    secret: ${vault:ALPACA_SECRET}

This is the part worth getting right. A leaked .env is the most common way hobby trading bots lose funds. Vault references keep the secret out of your repo, out of the prompt, and out of any logs the agent writes.

Step 3: Set the Schedule

The schedule is a cron expression on the deployment. For a 7:00 AM ET weekday brief:

0 7 * * 1-5

Set the timezone on the deployment so the cron fires in your market’s hours, not UTC. Common patterns:

Start with one job. A single working scheduled agent is worth more than five half-tested ones firing at once.

Step 4: Test Before You Trust the Schedule

Do not wait for the cron to fire to find out the task is broken. Trigger the deployment manually first. The console has a run-now button that starts a session immediately with the same config the schedule would use.

Check three things on that first manual run:

1. The connector authenticated. If the vault reference is wrong, this fails first.
2. The output landed where you expect (the email arrived, the file wrote).
3. The agent stopped after the task instead of looping or asking a question. A scheduled run has no human to answer prompts.

Once a manual run is clean, enable the schedule and check the next real run’s logs.

Step 5: Add an Execution Agent (Carefully)

A research brief is low risk. A scheduled agent that places orders is not. If you want a daily DCA or rebalance agent, keep it on a tight leash:

• Give it a hard cap per run (max dollars, max one order).
• Have it write every decision to a log before acting.
• Run it on testnet or paper first for at least two weeks.
• Keep its vault keys scoped to a sub-account, not your main account.

A DCA agent’s task might read: “Buy $50 of BTC at market. Do this exactly once and stop. If the order does not confirm within 60 seconds, log the failure and do nothing else.” The narrower the instruction, the less a scheduling or model error can cost you. The AI DCA bot guide covers the sizing logic in more depth.

Managed Agents vs Local Cron

Both still work. The choice comes down to what you are running and how much you want to maintain.

If you already have a stable local setup that you trust, there is no rush to move it. The vault and managed uptime are the real reasons to switch: you stop babysitting a server and stop storing keys in plaintext.

Bottom Line

The June 2026 Managed Agents beta makes scheduled Claude trading workflows a console setting instead of a server you maintain. Start with a research-only brief, store every secret in the vault, test with a manual run, and only graduate to order-placing agents once a paper version has run clean for a couple of weeks. Keep research and execution as separate agents so a bad schedule can never trade by accident.