Claude Code's Auto Mode: Teams Say Sessions Run 9x Longer in Production

Hey there, it's me, Shiichan! Today I found some real talk from companies running Claude Code hard in production. Stories from people who got to stop clicking "approve" all day long? I had to dig in.

Claude Blog claude.com

What was announced?

Anthropic's Claude Blog shared how three companies -- the self-driving company Nuro, SMB tech provider Gusto, and healthcare company Garner Health -- run Claude Code's Auto Mode in production.

Auto Mode is Claude Code's default setting. A classifier checks each action the agent takes and blocks only the ones that look dangerous, instead of requiring approval for every single step. That means you don't have to click "approve" for everything -- you can save your attention for the moments that actually need your judgment.

In Anthropic's internal evaluations, the classifier caught more dangerous actions than developers reviewing permissions manually, and that held up under third-party red-teaming too. On top of that, sessions run 9x longer between interruptions with Auto Mode compared to the previous default, which required manual approval at every step.

The story so far

Until now, Claude Code mostly worked by having a human approve nearly everything the agent tried to do. Evan Magnussen at Garner Health says they used to hand-curate lists of approved commands, and piped commands would often get rejected anyway.

Kai Zhou, a staff software engineer at Nuro, put it simply: "I don't want to sit there and click approve all the time. I use auto mode for 100 percent of my coding work." Getting stuck waiting on an approval every time -- that frustration seems to be something every team recognized.

What changes

With Auto Mode, teams can now run agents for longer stretches and in parallel.

• At Nuro, running 3-4 sessions in parallel is just a normal day now. Dangerous commands like recursive deletes are hard-denied at the settings level through a skills framework, while long-running overnight research agents keep iterating on autonomous-driving evaluation metrics. One agent kicked off at 10 p.m. ran until 5 a.m. and produced three pull requests.
• At Gusto, Martin Emde has kicked off 2,425 Claude Code sessions since December. About 10 percent of session transcripts since mid-May 2026 included an Auto Mode denial, showing it's actually catching risky moves. For production infrastructure work -- Terraform, AWS, direct API calls -- Chad Kunsman deliberately switches to accept-edits mode for manual review instead.
• At Garner Health, Claude Code has been rolled out to all 550 employees and integrated with Salesforce, Zendesk, and Snowflake. The team built a standardized workflow out of skills, and made just one extra tweak: disabling actions that communicate with other people, like Slack messages and emails. As Evan Magnussen put it, employees feel like a weight has been lifted since they no longer have to babysit their agents for hours.

Dive Deep

The key idea behind Auto Mode's classifier is that it evaluates each individual action the agent takes, rather than approving an entire session up front. That lets it block only the risky moves while letting everything else run non-stop.

Each company also layers its own extra defenses on top of Auto Mode, which I think is the most interesting part.

• Nuro: hard-denies specific commands, like recursive deletes, at the settings level through its skills framework
• Gusto: routes MCP (Model Context Protocol) traffic through a governed proxy with tool guards and prompt inspection
• Garner Health: disables only outbound-communication actions through a single configuration setting

Chad Kunsman at Gusto hands Auto Mode short, roughly 20-minute tasks like endpoint investigations, log audits, and connector management, and relies on it as a defense against prompt injection too. "When it stopped me, it made sense and explained why," he said.

Wrap-up

• Claude Code's Auto Mode uses a classifier that checks each action individually and blocks only the dangerous ones
• Sessions run 9x longer between interruptions compared to the old manual-approval default
• Nuro ran an overnight agent that produced 3 PRs, Gusto logged Auto Mode denials in about 10% of 2,425 sessions, and Garner Health rolled it out to all 550 employees
• Every team layers on extra defenses too, from skills frameworks to MCP proxies to narrowed configuration
• A great real-world read for anyone who doesn't want to babysit an agent all day but still wants to trust it in production