Deepagents Code Stops Asking for Permission — Auto Mode Is Now Stable

Deepagents Code shipped two updates on July 24, 2026 — version 0.1.46 and version 0.1.47. The big news: Auto mode is now generally available. If you use Deepagents Code as your terminal coding agent, this changes how much you interact with it during a session. Instead of approving every file edit and terminal command one by one, Auto mode lets the agent decide which actions are safe enough to run without asking.

What Auto mode actually does

Every coding agent faces the same tension. On one side, you want the agent to work fast — make the edits, run the tests, fix the bugs, move on. On the other side, you want control. You do not want an AI tool deleting files, running destructive commands, or pushing code without your knowledge.

Auto mode is Deepagents Code’s answer to that tension. When Auto mode is off, the agent asks you to approve every action before it runs. That is safe but slow — a complex refactor might require fifty approval clicks. When Auto mode is on, the agent evaluates each action against your active goal and rubric (the instructions you gave it about what it should and should not do). Actions that align with your goals get authorized automatically. Actions that seem risky still prompt you.

Think of it like this: you are supervising a junior developer. In manual mode, you review every pull request line by line. In Auto mode, you tell them “here are the guidelines, go ahead and build it, but check with me if anything looks off.” The junior developer still follows your standards — they just do not interrupt you for every routine decision.

What changed in version 0.1.46

The 0.1.46 release made several improvements to Auto mode:

Goal-criteria acceptance is now configurable. Before this release, Auto mode used a fixed set of criteria to decide whether an action was safe. Now you can configure how strictly the agent evaluates actions against your goals. If you are working on a low-risk project and want the agent to move fast, you can loosen the criteria. If you are working on a production codebase with strict requirements, you can tighten them.

Actions from active goal and rubric directives are auto-authorized. This sounds technical, but it means something practical: if your goal says “refactor the authentication module,” and the agent edits an authentication file, that edit gets approved automatically. The agent does not need to ask you “is it okay to edit this file?” when you literally told it to do that. This eliminates a class of annoying approval prompts that waste your time.

Redundant approval prompts are eliminated. Before 0.1.46, Auto mode sometimes asked for approval on actions it had already authorized. That defeats the purpose. The fix removes these duplicate prompts so the agent flows through your task without unnecessary interruptions.

Classifier failure logging. Auto mode uses a classifier to evaluate whether actions are safe. When that classifier fails (timeout, error, etc.), the failure is now logged instead of silently falling through. This means you can diagnose why Auto mode might be behaving unexpectedly.

Recursion limit raised to 2000. Deepagents Code had a default recursion limit that could cause the agent to stop mid-task on complex, multi-step projects. The new limit of 2000 is configurable and gives the agent room to handle larger refactors without hitting a wall.

What changed in version 0.1.47

The 0.1.47 release is smaller but adds one feature that power users will appreciate:

Yolo mode in the Shift+Tab approval cycle. When you are in manual mode (Auto mode off), you still approve actions one at a time. Pressing Shift+Tab cycles through approval modes. Version 0.1.47 adds “yolo” as one of those modes. When yolo is active, every approval prompt is automatically accepted — no clicks needed.

The difference between Auto mode and yolo: Auto mode evaluates each action against your goals and rubric before approving. Yolo does not evaluate — it approves everything. Auto mode is supervised autonomy. Yolo is full trust.

Use Auto mode when you want the agent to be smart about what it approves. Use yolo when you know the agent is about to do a series of safe operations and you do not want to babysit it.

Changelog link before update prompt. When a new version is available, the agent now shows you the changelog link before asking if you want to update. This is a small quality-of-life change — you can read what changed before deciding whether to upgrade.

Claude Opus 5 support

Both versions added Claude Opus 5 as a supported model. Opus 5 is Anthropic’s latest frontier model with a 1M token context window. If you are using Deepagents Code with Anthropic models, you can now select Opus 5 as your model backend. The 1M context window means the agent can work with very large codebases without losing track of earlier context.

Hooks v2 capability snapshots and session transcripts

Version 0.1.46 also hardened the Hooks v2 system. Hooks are lifecycle callbacks that run at specific points in the agent’s execution — for example, before a tool call or after a file edit. The new capability snapshots let you see exactly what tools and capabilities are available at each hook point. Session transcripts now record hook execution, giving you a complete audit trail of what happened during a session.

For teams using Deepagents Code in regulated environments or with strict code review processes, this audit trail is valuable. You can trace every action the agent took, every hook that fired, and every approval decision that was made.

How to update

If you have Deepagents Code installed, run the install script again to get the latest version:

curl -LsSf https://langch.in/dcode | bash

Or if you installed via pip:

pip install --upgrade deepagents-code

To enable Auto mode, use the Shift+Tab cycle or the configuration settings. To configure goal-criteria acceptance, check the Deepagents Code documentation at docs.langchain.com/deepagents-code.

Why this matters

Auto mode going stable is the kind of change that sounds incremental but changes daily workflow. Every coding agent — Claude Code, Hermes Agent, Gitlawb Zero, OpenCode — deals with the same approval friction. Agents that require constant human approval feel slow. Agents that skip approvals entirely feel risky. Auto mode sits in the middle: it respects your goals and rubric while reducing the number of times it interrupts you.

For developers who tried Deepagents Code before and found the approval prompts too frequent, the stable Auto mode is worth a second look. For existing users, the configurable goal-criteria acceptance and yolo shortcut give you more control over how much supervision the agent needs.

The coding agent space moves fast. Deepagents Code is not the most popular terminal agent, but with Auto mode now stable and Claude Opus 5 support, it has a feature set that competes with the best tools in the category.