Default approval mode for goose

[zanesq]

What do people want for the default approval mode, auto or smart or manual approve?

#2806

[DOsinga]

FWIW I like living on the edge and think autonomous is best.

either way, we should make the setting such that this is an easy thing to change, rather than on the second page

[zanesq]

Interesting idea, like part of onboarding/setting up providers?

[EbonyLouis]

yea I like this idea too it needs to be obvious to the user that they can change this and apart of the onboarding sounds good

[joeeuston-dev]

I really like autonomous as well. Kinda because I really want to see just how fast goose can go.

[The-Best-Codes]

I personally prefer manual approval (thought smart approval, where an AI model decides, is also nice) in AI tools.
I've seen some AI models run some really dumb or destructive commands that I would never run myself, so I always like to be able to reject commands and leave Goose unattended without worrying too much.

[ewann]

+1

Ideally: conservative, manual mode as a default. But really easy & well signposted for users to switch to one of the others. IMO it would even be acceptable to nag when not in autonomous, if there is a env var to suppress the nag if you like living with lower-case-c-conservatives

[EbonyLouis]

I think the default should be autonomous mode, so new users get the full Goose experience right away. From there, they can always switch depending on their preference.

[angiejones]

autonomous - it's an agent! let it be freeeee

[DOsinga]

seems like we all want this is to be easier to configure. you can already do this from the bottom of the chat bar, easy enough. what if that was only for the current conversation though? I think that would make sense to me

[ynniv]

I also like to live dangerously, and vote for autonomous be the default

[ynniv]

That's not true. It's rare for anonymous to be dangerous, and people are very reluctant to enable it. Everyone will likely be better off with autonomous as a default that's easy to change

[ewann]

That's not true. It's rare for anonymous to be dangerous, and people are very reluctant to enable it. Everyone will likely be better off with autonomous as a default that's easy to change

Within a few days of starting to experiment with it, I’ve already had it delete a bunch of env vars without having correctly understood it should have moved those values elsewhere. I’m inclined to say this was model / provider specific, adding some support to #4324 (comment)

[taniandjerry]

By default: autonomous mode for sure. I occasionally need manual mode for specific tasks, and will toggle it on when I need to.

It could be neat to have a quick keyboard shortcut to toggle between modes while goose is running, or to tell goose 'use manual mode for this next task' without impacting the global default.

[ynniv]

I would love a way to raise my hand in the CLI. A kind of "hold up", without a hard interrupt

[blackgirlbytes]

Autonomous..but maybe during the sign up / onboarding flow..we can ask the user what they prefer..if theyre not used to agents..it can be a jarring experience LOL. As Ive gotten used to auto mode, there's no turning back for me.

[jmdaly]

I personally use autonomous mode all the time, but I wonder about the risks for new users if that's the default. It'd be unfortunate if a new user tried Goose, unknowingly ended up in an autonomous mode, and the LLM called a tool that damaged their local system / work. Just my 12 cents :)

[The-Best-Codes]

That's close to how I see it. I'd be paranoid that new users (at least developers) would think, “what a security nightmare!” if it turns on autonomous mode by default.
And from a security perspective, Goose should follow the "Principle of Least Privilege" rule, where only the minimum permissions needed for the tool to operate are enabled by default and others are opt-in.

[ynniv]

Maybe fully autonomous IS the least sufficient privilege 😎. I get it ... I've worked at fintechs and written highly sensitive code. But defaults matter, and based on running goose fast and loose straight in bash for ~ 16 months, it's surprisingly safe. Using anything other than autonomous is leaving a lot on the table.

Maybe it should only be the default for known good models though? I've been using Claude Sonnet this whole time

[ynniv]

My go to anecdote is asking John Carmack for help with your code. Are you going to micromanage John?

[The-Best-Codes]

Maybe fully autonomous IS the least sufficient privilege

LOL. But seriously, from a security POV, the least sufficient privilege would be the ability for the code to make a network request to the AI model's API. The next thing would be to read files, then apply diffs to files, then write files, etc.

My go to anecdote is asking John Carmack for help with your code. Are you going to micromanage John?

Hmm, this is a good example. LLMs like Claude 4 do write the best code when they have full permissions to do whatever they need, just like a developer. But with other LLMs, it's not always a great idea. I'm especially thinking of Gemini 2.5 Pro, which can get in a panic when it messes things up and just try removing the whole codebase.
I like @iandouglas's idea, having a per-model permission system would be really nice, while still being able to override it per-session, of course.
But even in that case, there still has to be a default permission setting (like maybe enable autonomous mode by default for Claude 4 and some other good models, but disable it on others).

[iandouglas]

I'd love to configure it on an LLM basis to be honest ... there are some LLMs that I know cannot do tool calling, so might as well just park those in 'chat' mode, where something with tooling could be autonomous if i trust it through experience, or put it in approval mode if I'm trying something new and don't know how well it's gonna work etc.

[ewann]

I've been trying not to comment here again 😅 but I couldn't let this one pass:

User:
how many instances of goose desktop are currently running on this machine?

Agent:
<snip>
To address this, we can try to identify which processes are associated with Goose Desktop and terminate any unnecessary ones
<snip>

I'm certainly in favour of on an LLM basis 😅

[cruftex]

The default would be important for new users, not an existing community what already has an understanding how the tool operates. A default isn't relevant at all for experienced users, since the can always switch to expert setting. So the decision making is a bit flawed.

It still should be documented that the agent is allowed to issue any shell commands on your system.

So if you like to get better adoption and not frustrate new users please think of them, rather then what you want. You have the experience and the trust in the tool that you build. New users don't.

What would a new user expect?

What would a new user need to know?

Also maybe don't call it auto approval, if no approval happens, but better skip approval?