Does a Programmer Need an 'Uninsured' Button for an AI Agent?

Does a Programmer Need an 'Uninsured' Button for an AI Agent?

Anthropic recently added an environment variable to Claude Code that removes the final safety barrier against dangerous commands like file deletion. It is named without a hint of irony — as long and literal as possible. I have long been tracking how coding agents are evolving from cautious assistants to full-fledged autonomous executors, and this specific step made me reconsider my own approach to working with such tools. I am sharing my column-guide: what has changed with Claude Code and Codex over the last few weeks and how I prioritize speed versus control.

1. Why remove the brakes on an agent at all?

The old flag --dangerously-skip-permissions already allowed Claude Code to work without constant confirmations. The new variable goes further:. My position here is unambiguous: such a mode only makes sense in isolated containers and disposable environments where losing files is not a tragedy. On a live project with a production branch, I wouldn't touch such a button even under deadline pressure.

2. Autonomy by default is not an option, but the norm

In parallel with the radical flag, Anthropic has taken a step toward 'soft' autonomy:. This is a fundamentally different philosophy — the agent decides for itself when it can act without pausing, but it retains built-in 'common sense' by default. In practice, this is most convenient for routine tasks: refactoring, writing tests, and minor style edits. For such things, I trust automation — it really saves time that used to be spent on endless 'yes, continue' prompts.

3. Projects — supervision instead of manual tab switching

Another change that I like more than the others:. Previously, the parallel work of several agents turned into a chaos of open terminals. Now it is a managed process with a unified panel — something that was sorely lacking for any task larger than a single file.

4. Code Review is now generally available — and it changes the workflow

I would also like to note that the code review feature, which catches bugs before production, has left beta:. In my opinion, it is review — not code generation — that is now becoming the main battlefield between agents. AI is already quite good at writing code, but catching its own mistakes before they reach the main branch is a task of a different order of complexity.

5. OpenAI has its own race: models for Codex are updated almost weekly

While Anthropic balances speed and safety, OpenAI is expanding its model lineup for Codex. The fresh model has received notable positioning regarding usage economics:. This is a convenient benchmark for choosing a model for a task: fast and cheap models for routine work, and 'heavier' ones for architectural decisions and complex bugs.

My conclusion

Tools like Claude Code and Codex are maturing rapidly: from simple prompts to agents that decide for themselves when to act without confirmation. This saves a lot of time, but the responsibility for the consequences does not go anywhere — it simply shifts to the person who flipped the switch. I would advise keeping autonomous modes enabled on isolated branches and in test environments, while leaving any destructive operations on production code under manual control, regardless of how 'smart' the agent appears.

If you want to try working with modern models for programming and other tasks without unnecessary infrastructure setup, check out the GEMERA AI tool catalog: it contains current models for code, text, and generation in one place.