In May 2026, Anthropic engineers shipped around 2,500 code changes a day. By September, they passed 30,000 edits a day—a pace no human can track by eye. In her talk, Claude Code engineer Lydia Hallie explained how the team scaled development more than tenfold, and Nate Herk broke down the config details.
Earlier in the year, working with Opus 4.5 felt like supervising an intern: you'd kick off a task, keep correcting the model, and pack CLAUDE.md with all-caps shouts like IMPORTANT and NEVER. After upgrading to Opus 5.5, the team found that stern rule lists actually confused the stronger model and triggered pointless tool calls. They trimmed the file down to what isn't in the code itself—folder maps for wikis, brand guides, and finance docs. They didn't even have to clean it by hand: /doctor checks instructions against the model and strips out the clutter.
The Auto Mode Classifier Instead of Manual Approvals
That raises an obvious question: if the model roams through files on its own, who clicks "allow" before every terminal command? Before, you either clicked manually or disabled checks at your own risk. Now Claude Code enables Auto Mode by default: a separate classifier inspects commands on the fly and only pauses sketchy ones, like pushing to a deploy repo outside the current folder.
In Anthropic's August test, humans caught only 14% of clearly dangerous commands on average, while Auto Mode blocked 89%. That's a huge jump, but if one in ten risky commands still slips past, you can't rely on a classifier alone. The takeaway is simple: let Auto Mode handle routine work, and block irreversible actions with hard rules before it even runs.
Hard Deny Rules, Sandboxing, and the exit 2 Rule
Lydia uses one simple test for what deserves a hard ban: can you undo the mistake? Bad code is easy to revert in Git. Deleted drafts, leaked .env keys, or a dropped database table are gone for good.
To cover those risks, teams add .claude/config.json to the repo. Say you ask an agent to wire up a Redis cache—you can block pushes to main, table drops, and secret exports up front, while attaching a pre-commit test script:
{
"autoMode": {
"enabled": true,
"allowedTools": ["FileRead", "FileEdit", "DirectoryList", "Bash"]
},
"denyRules": [
"git push --force*",
"git push origin main",
"rm -rf /",
"drop table*",
"export AWS_SECRET_ACCESS_KEY*"
],
"hooks": {
"preCommit": "./scripts/verify-and-test.sh"
}
}
These filters run before the Auto Mode classifier. If the agent tries to dump production keys or push to main while debugging Redis, the rule stops it without bothering you. Keep in mind that deny rules only inspect the command string, not what third-party test scripts do underneath. To stop leaks via packages, teams turn on /sandbox, which locks down file and network access at the OS level. For long runs, Lydia moves tasks to cloud sessions on disposable machines with no local keys, where the repo config loads automatically.
Next, you need to make sure a multi-hour run didn't break app logic. Even on a straightforward auth migration, one prompt can leave the old login active or log everyone out. Hooks catch this, but there's a catch: standard linters exit with code 1 on failure, and in several environments Claude Code treats 1 as a finished step. To send the agent back to fix the bug, your script must return exit code 2.
#!/bin/bash
npm run lint || { echo "Linting failed!"; exit 2; }
npm test || { echo "Tests failed! Returning control to Claude."; exit 2; }
exit 0
As soon as a check fails with exit 2, the model sees the error and fixes the mocks or code until it gets a clean exit 0. Once a PR is ready, a separate Claude session with fresh context reviews it first, and a second group of agents double-checks for bugs. Humans only review high-risk changes—though at 30,000 edits a day you need Squash & Merge to keep Git readable, and auto-merging docs still requires a human eye on legal and public API details.
One Orchestrator Instead of Twelve Tabs
Even spinning up separate sessions is moving to a coordinator in Claude Code Projects, now in beta on Pro Max and Ultra plans. You hand one main session your weekly task list, and it spawns worker threads for each feature. You can put a strong planner like Fable 5.2 in the coordinator seat and assign the coding branches to Opus 5.5.

