In May 2026, InfoQ published Inside Claude Code Auto Mode: Anthropic’s Autonomous Coding System with Human Approval Gates, drawing intense attention to operational models combining Claude Code's long-running autonomous execution in "Auto Mode" with human approval gates. The approach of progressing hundreds of steps overnight while always summoning a human before dangerous actions provides a clear answer to the question of "how much scope to entrust to AI" that has long challenged custom development teams.
This article outlines how to safely operate Auto Mode in custom development, covering human approval gate design, diff review workflows, and the assembly of regression detection.
What Auto Mode changed
Previously, Claude Code predominantly used an interactive model: "advance a few steps → human checks → next few steps." The biggest shift with Auto Mode is that "the agent chooses when to summon a human." Comparing InfoQ's analysis with Anthropic's official documentation highlights three primary operational pillars.
| Dimension | Traditional mode | Auto Mode |
|---|---|---|
| Initiative in progress | Humans provide the next instruction | The agent advances work and requests approval at milestones |
| Handling dangerous operations | Prompted every time | Prompted only when matching a predefined "critical operations list" |
| Audit logs | Conversation history only | Structured persistence of decision logs at each step |
| Parallelism | One task per session | Premise of running multiple sessions simultaneously |
In particular, pre-agreeing on a "critical operations list" on a per-contract and per-project basis is a significant step forward, as it allows putting what was previously ambiguous in Cursor or GitHub Copilot CLI—namely, "how far we can entrust tasks to AI"—into a form that can be written into contracts.
Three tiers of approval gates to establish in custom development
When operating Auto Mode in custom development, approval gates are designed in three tiers: "Automatic Approval," "Synchronous Approval," and "Asynchronous Approval."
[Tier 1: 自動許可（Auto Allow）]
├ 読み取り操作（git status, ls, grep, read_file）
├ ローカルテスト実行（npm test, pytest）
├ ドラフトブランチへの commit
└ Pull Request の draft 化
[Tier 2: 同期承認（Sync Approval）]
├ npm install / 依存追加
├ DB マイグレーション（dev 環境）
├ main / develop へのマージ提案
└ 環境変数の参照
[Tier 3: 非同期承認 + 二重確認（Async + Double Check）]
├ 本番 DB への接続
├ シークレットの読み取り（.env, GitHub Secrets）
├ 外部 API への送信（Slack 投稿、メール送信、決済 API）
└ git push --force / branch 削除Tier 3 operations are designed so that "progress cannot be made without multiple people present," operating such that execution only occurs after reactions from two or more authorized approvers are gathered in Slack or Teams. This is a theme that should be built in tandem with the philosophy of "preventing AI from directly executing destructive operations" discussed in Production DB Deletion Guardrails for AI Agents.
Template for "Auto Mode terms of use" to include in custom development contracts
To use Auto Mode safely in custom development, the greatest preventive measure is to reach an agreement with the client on the scope of use during the contract phase. The key points to cover as contract clauses are shown below.
| Clause | Details | What the client should verify |
|---|---|---|
| Target environments | dev / staging only. Production is manual only | Explicit statement that "AI does not touch production directly" |
| Critical operations list | Defined in an appendix for each project | Whether the list is exhaustive and accurate |
| Audit log retention | Retain decision logs for all sessions for 1 year | Log ownership and access rights |
| Model version pinning | Pin the model during the project; give prior notice for changes | Verification period upon switching |
| Demarcation of responsibilities | Explicitly state that the contractor bears responsibility for erroneous code produced by AI | Handling when the client operates the AI directly |
| Exclusion from training | Do not include client code in training data | Verification of API plans (e.g., via Bedrock / Vertex) |
Model version pinning is especially critical in practice, as a minor update to Claude in the middle of a project can change code generation tendencies and cause regressions. We incorporate a flow into the contract to pin anthropic-version and verify compatibility in a verification environment at the beginning of each month.
Diff review mechanism: Visualizing the volume written by AI
When running Auto Mode, automated commits on the scale of thousands of lines become routine. To guarantee quality in custom development, a mechanism that visualizes the ratio between the volume written by AI and the volume reviewed by humans is essential.
Here are the four metrics you should maintain on your project dashboard:
- AI automated commit rate = Number of commits pushed directly by AI / Total number of commits
- Human review rate = Lines in PRs approved by humans / Total lines added
- CI pass rate = First-time CI pass rate among PRs submitted by AI
- Regression rate = Number of bugs originating from AI-derived commits / Number of AI-derived commits
If the human review rate drops too low, accountability to the client cannot be maintained, so it is safest to make it a baseline rule to prohibit automatic merging in Auto Mode and require a human to click the button for all PRs. This concept is directly continuous with the operation of "overnight batch AI" discussed in Claude Code 2026 Latest Feature Summary and Maintenance for Clients with Vercel Open Agents.
Regression detection: The final defense of custom development quality
As Auto Mode becomes more capable, code beyond what human eyes can fully track enters the repository daily. Unless you set up multiple layers of regression detection, the scenario of "code that no one remembers" breaking six months later and triggering inquiries from the client becomes a reality.
| Layer | Tool examples | Detection target |
|---|---|---|
| Static analysis | semgrep / eslint / typescript —strict | Syntax, types, known anti-patterns |
| Unit tests | Vitest / Jest / pytest | Function-level behavior |
| Integration tests | Playwright / Cypress | Entire UI flows |
| Visual regression | Chromatic / Percy | Screen diffs |
| Performance regression | Lighthouse CI / k6 | Response times |
| Security regression | trivy / gitleaks / semgrep —rules security | Introduction of vulnerabilities |
Especially in projects where AI touches 50 files overnight, visual regression and performance regression are crucially important. Even if only npm run test passes, UI breakages and page slowdowns tend to be noticed only when the client interacts with the system themselves. This is a theme that must be built in conjunction with the "testing strategy in the AI era" discussed in Vitest 4.1 AI Agent Reporter.
Beware of misjudging cost structures
While Auto Mode accelerates development speed, the fact that operational costs for human approvals, regression detection, and contract preparation increase should be shared with clients during the estimation phase.
The fact that introducing Auto Mode does not simply reduce costs is a key point to explain to clients. Agreeing upfront that "delivery is faster, but costs for verification and contract preparation increase" prevents clients from later complaining that "it wasn't as inexpensive as expected." Because the required team structure varies greatly depending on the number of authorized approvers, whether overnight operations run, and to what layers regression detection is deployed, organizational design and costs must be hammered out on a per-project basis.
Four common pitfalls
Finally, here are common pitfalls that teams often encounter when operating Claude Code Auto Mode in custom development.
Pitfall 1: Leaving the "critical operations list" outdated
As development progresses, new external APIs and database tables are added, causing the initially determined list to diverge from reality. Always perform a list review every sprint.
Pitfall 2: Audit logs disappearing due to storage limits
Auto Mode generates several megabytes of logs per session. Operating with default settings leads to incidents where logs no longer exist during incident investigations six months later. Build in daily archiving to S3 / R2 from the outset.
Pitfall 3: "Trials" of model updates leaking into production
When a new model is released, it is tempting to try it out, but switching models mid-project introduces regressions. Set up separate validation branches and pin the model for production projects.
Pitfall 4: The temptation for clients to operate the AI directly
Client stakeholders sometimes request that they want to try Claude Code themselves. Since allowing direct access breaks the demarcation of responsibilities, make it a rule that "execution is routed through the contractor's environment, and all logs are shared."
Summary: Balancing speed with safeguarding client assets
Claude Code Auto Mode is a breakthrough mechanism that delivers long-duration autonomous execution at realistic costs, but using it in custom development requires preparing the four-part foundation of contracts, approvals, auditing, and regression detection upfront.
Whether Auto Mode can be introduced into custom development ultimately hinges on what you place in Tier 3. If a project allows production databases and secrets to remain under human control, overnight execution is a viable option; if environments are not even isolated to begin with, then quite simply, there are issues to resolve before considering Auto Mode.
It is entirely fine if you are still at a stage where you cannot tell which side your project falls on. The outline of the required structure will begin to emerge simply by comparing your existing approval workflows against a draft list of critical operations. Because subsequent design depends entirely on your project's environment configuration and contract structure, please start by telling us about your current situation via our contact form.
