Love this! AI agents can be so abstract to many people and this project really makes it feel much more approachable. Makes me think of Game Dev Story! Would be awesome to see little thinking bubbles over them to show what they are doing. Ultimately I see promise in making it easier to visually see what is happening in the system.
Completely agree, we found it helped explaining it to our non-technical friends as well.
We do have thinking bubbles but they only show up based on the task the agent is doing. Perhaps we'll add a toggle or something to give people the option to have them always on.
Maybe a bit of an odd one, but can you decorate the office? I'm wondering like have you abstracted the decor elements into something that is straightforward to extend? How easily could I give my office a new espresso machine or something?
This is a fun question. The answer is yes, you can redecorate and move things around.
Right now it's all built in phaser, and furniture is pretty straight forward to build and deploy. But it requires modifying the source. We want to add support for easy drop in decoration in future updates.
iMessage is the one that changes the category for me. "Agents that can text real people" is a different thing than a multi-agent demo. I've been hacking on something similar and the hardest problem wasn't orchestration it was figuring out when to bother a human vs. just deciding. How does auto-approval work, is it all-or-nothing or per task?
There are two levels of auto approve, first level is auto-edit which is basic read and write, and basic bash tools (these can be configured to be any arbitrary bash command).
The second level is called auto approve and is for more complex bash commands. Generally the model will ask permission before running one of these big commands, but you can allow all. Right now, it's global across the instance, but we're working on making it more granular.
Also, there is a deny list of certain commands which you can customize to prevent bad behavior (like rm -rf, etc...)
We want to wire the approval process to imessage or whatever channel, but we need to first auth the imessage session to make sure it's coming through from the owner and not someone else communicating through the same channel.
Love the pixel office. Such a fun way to make multi-agent work less abstract. Being able to actually watch agents walk to their desks and pick up tasks makes it way easier to follow what's happening than staring at terminal logs. Curious if the orchestrator handles cases where two agents need to edit the same file.
Thank you James, great question! Generally, when the orchestrator assigns parallel work it does it in groups, with everyone in the group working on different files. When the group is done it passes the work to the next group which can then edit those same files.
We're working on simultaneous editing of the same files using git, but we want to ensure changes are merged in an intelligent way.
Not allowed? This is easy to find in the Agent SDK docs
"Unless previously approved, Anthropic does not allow third party developers to offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK. Please use the API key authentication methods described in this document instead."
We saw this, but thought it was for Agents calling the API directly. Outworked is just a wrapper around your CLI using the existing agents and sub-agents in your Claude Code installation.
It's a great point though and we'll need to read into this more in-depth. Appreciate you raising this.
Right, pretty sure people got their accounts shut down for doing this kinda thing in the early days of OpenClaw before it was renamed and when it supported login. I guess you can take your chances but API is probably the only safe way to use this.
"animal crossing-style" is a bit of a stretch
Yeah, I was expecting something like the Animal Crossing dialog bubble or something. At least put Tom Nook as the boss character.
This is such a good idea. A nondescript non-copyright infringing raccoon character as "The boss" would be perfect.
I guess the AI didn't know what Animal Crossing looks like. Cool visualization though.
yeah.. maybe should have said "inspired by"
Not even. Maybe briefly saw an advertisement, but didn't click it.
tbh just looked at the cover you know?
Love this! AI agents can be so abstract to many people and this project really makes it feel much more approachable. Makes me think of Game Dev Story! Would be awesome to see little thinking bubbles over them to show what they are doing. Ultimately I see promise in making it easier to visually see what is happening in the system.
As a visual-oriented person I can completly agree. Being able to visualze makes things make so much more sense.
Completely agree, we found it helped explaining it to our non-technical friends as well.
We do have thinking bubbles but they only show up based on the task the agent is doing. Perhaps we'll add a toggle or something to give people the option to have them always on.
Maybe a bit of an odd one, but can you decorate the office? I'm wondering like have you abstracted the decor elements into something that is straightforward to extend? How easily could I give my office a new espresso machine or something?
This is a fun question. The answer is yes, you can redecorate and move things around.
Right now it's all built in phaser, and furniture is pretty straight forward to build and deploy. But it requires modifying the source. We want to add support for easy drop in decoration in future updates.
The parallel agent coordination is what makes this interesting. Most agent wrappers are just single-agent loops with extra steps.
Yes, we actually added it because we were getting impatient with how long some implementations were taking. Appreciate this comment - thank you!
I like the visualization, but in terms of orchestration, how does it compare to CC’s built in agent swarms?
Great question! It actually takes advantage of the sub-agent swarms since it is directly connected to your Claude Code instance.
We want to implement agent teams as well, but it is still an experimental Claude Code feature. So it's more of a secondary priority right now.
iMessage is the one that changes the category for me. "Agents that can text real people" is a different thing than a multi-agent demo. I've been hacking on something similar and the hardest problem wasn't orchestration it was figuring out when to bother a human vs. just deciding. How does auto-approval work, is it all-or-nothing or per task?
There are two levels of auto approve, first level is auto-edit which is basic read and write, and basic bash tools (these can be configured to be any arbitrary bash command).
The second level is called auto approve and is for more complex bash commands. Generally the model will ask permission before running one of these big commands, but you can allow all. Right now, it's global across the instance, but we're working on making it more granular.
Also, there is a deny list of certain commands which you can customize to prevent bad behavior (like rm -rf, etc...)
We want to wire the approval process to imessage or whatever channel, but we need to first auth the imessage session to make sure it's coming through from the owner and not someone else communicating through the same channel.
Love the pixel office. Such a fun way to make multi-agent work less abstract. Being able to actually watch agents walk to their desks and pick up tasks makes it way easier to follow what's happening than staring at terminal logs. Curious if the orchestrator handles cases where two agents need to edit the same file.
Thank you James, great question! Generally, when the orchestrator assigns parallel work it does it in groups, with everyone in the group working on different files. When the group is done it passes the work to the next group which can then edit those same files.
We're working on simultaneous editing of the same files using git, but we want to ensure changes are merged in an intelligent way.
Are you planning a Linux release?
In theory it should build in Linux (electron) but we haven't tested it yet. Things like the imessage integration wouldn't work. Thanks for asking!
Haha this is really cool! I imagine it would be a nice tool to teach kids about working with agents.
This is a really neat idea, we hadn't thought about the education angle. Thanks for sharing!
does it use the claude code api or the claude code cli? You know, the claude code api is more expensive.
I also hope it can have a webapp version, rather than electron. because most of our work are on a remote server.
It actually uses the Claude code SDK so it plugs into whatever you already have.
It can use API/CLI or even if you have a private hosted instance.
We're actually working on a remote web app version but its a little trickier to wire up.
These are great questions - thank you!
If it uses the SDK then it's token burn? Or can it "legally" use your Claude.ai MAX account, your subscription account?
Great question - we're going to look into this in-depth.
Not allowed? This is easy to find in the Agent SDK docs
"Unless previously approved, Anthropic does not allow third party developers to offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK. Please use the API key authentication methods described in this document instead."
https://platform.claude.com/docs/en/agent-sdk/overview
We saw this, but thought it was for Agents calling the API directly. Outworked is just a wrapper around your CLI using the existing agents and sub-agents in your Claude Code installation.
It's a great point though and we'll need to read into this more in-depth. Appreciate you raising this.
Right, pretty sure people got their accounts shut down for doing this kinda thing in the early days of OpenClaw before it was renamed and when it supported login. I guess you can take your chances but API is probably the only safe way to use this.
this seems interesting. Will give it a spin this weekend.
Thank you - let us know if you have any feedback. Appreciate you trying it out.
Sorry, but there's not even a single hint of a visual relationship to Animal Crossing here. I believe this title is inaccurate.