claude-code - 💡(How to fix) Fix [BUG] Cowork: "Reached maximum number of turns (100)" breaks long-running browser automation projects

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Product: Claude Cowork (macOS desktop app) with Claude in Chrome extension Model: Sonnet 4.6 Plan: Max Date issue started: Recently — workflows that previously completed in a single run are now failing

Summary

A scheduled Cowork project that audits CRM leads via browser automation (screenscraping DealerCenter) is hitting a hard ceiling of "Reached maximum number of turns (100)" after processing only 4–5 records, then halting with a "Something went wrong" error and "Try again" / "Go back" buttons.

Workflow context

• Daily CRM lead audit across two dealership locations (Omaha and Lincoln)
• Typical daily volume: 50–90 leads (today: 28 Lincoln + 58 Omaha = 86 leads)
• Workflow requires screenscraping because DealerCenter does not expose an API/MCP
• Each lead audit involves multiple page navigations, screenshots, and reads — easily 15–25 turns per lead

Impact

• This project was previously completing in a single unattended run
• It now requires the user to actively monitor and restart, processing only ~5 records per attempt before hitting the 100-turn ceiling
• To complete a daily audit under the new ceiling would require ~18 separately scheduled batches running over 4+ hours
• This is a fundamental regression for production browser-automation workflows

Questions for the team

1. Is the 100-turn ceiling new, or has it always been there and just recently started being hit?
2. Is the ceiling configurable per-project (similar to the SDK's --max-turns parameter)?
3. Is there a recommended pattern for long-running browser-automation tasks that need to process 50+ records in a single unattended run?

Reproduction

1. Set up a Cowork scheduled project using Claude in Chrome to navigate a CRM and process records one by one
2. Configure the workflow to handle 5+ records that each require multiple page interactions
3. Run the project — it will halt at "Reached maximum number of turns (100)" partway through

Screenshot

[paste your original screenshot of the error here using the image upload button in the editor]

What Should Happen?

Cowork should be able to complete long-running browser automation tasks in a single unattended run, the way it did previously. For a typical CRM lead audit workflow processing 50–90 leads, the project should run to completion without halting mid-task and requiring user intervention to restart.

If a turn ceiling must exist for safety/cost reasons, it should be:

1. Configurable per-project (similar to the SDK's --max-turns parameter), or
2. High enough to support real production workflows involving 50+ browser-scraped records, or
3. Replaced with an automatic continuation mechanism so the user isn't forced to babysit and manually restart.

Error Messages/Logs

Reached maximum number of turns (100)

You can restart the conversation from an earlier message.

Steps to Reproduce

1. Set up a Cowork scheduled project that uses Claude in Chrome to interact with a web-based CRM (in my case, DealerCenter — but this would apply to any browser-automation workflow involving many records).

2. Configure the project to audit/process a list of records one-by-one, where each record requires multiple page interactions (navigate to record, screenshot, read fields, navigate back, repeat).

3. Set the daily record count to a realistic production volume — 50 to 90 records.

4. Run the project (either scheduled or manually triggered).

5. Observe: Cowork halts after approximately 4–5 records with the error "Reached maximum number of turns (100)" and presents "Go back" / "Try again" buttons. Restarting only processes another ~5 records before hitting the same ceiling again.

Note: This is not a Claude Code CLI bug — it is a Claude Cowork (macOS desktop app) bug. Filing here because there is no separate public repo for Cowork issues. Please route to the Cowork team.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

Unable to specify exact version. Cowork doesn't expose a public version number to users that I'm aware of. The workflow was working as recently as [fill in the most recent date you remember it completing in a single run — last week? Two weeks ago? Best guess is fine].

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

IMPORTANT: This is a Claude Cowork issue, not a Claude Code CLI issue

Filing in this repo because there is no separate public repo for Cowork issues. The "Which terminal" field above does not apply — I am not using Claude Code from a terminal. I am using:

• Product: Claude Cowork (macOS desktop app)
• Browser agent: Claude in Chrome extension
• Operating system: macOS [your version]
• Plan: Max
• Model: Sonnet 4.6

Please route this to the Cowork team rather than the Code CLI team.

Business impact

I run a used car dealership and use Cowork daily for production operational workflows including:

• Daily CRM lead audits across two locations (50–90 leads/day)
• Daily ACH deposit reconciliation (Gmail → DealerCenter → QuickBooks)
• Google Business Profile review monitoring
• Facebook Messenger lead response automation

Multiple workflows are now hitting the 100-turn ceiling, requiring me to either babysit and manually restart, or split each workflow into many small scheduled batches. This is a significant degradation of the product's value for real-world business automation.

Related

I have a separate open issue regarding Cowork browser/JavaScript execution permissions not persisting. That issue is unrelated to this turn-limit bug but shares the same product surface.

Screenshot

[Drag and drop your error screenshot from this morning into the editor here]