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.
Right now purely file level. But the dependency analysis is really intelligent and we'll figure out what that would look like.
The long-tail point is true. We don't do any tracking in our implementation, but we've been trying hard to refine our total our total permissions approach to think about more edge cases such as this, while not being too annoying. We think this is a general tricky issue with AI alignment ('do what I mean not what I say').
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.
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.
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 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.
We're working on simultaneous editing of the same files using git, but we want to ensure changes are merged in an intelligent way.
I also hope it can have a webapp version, rather than electron. because most of our work are on a remote server.
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!
"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
It's a great point though and we'll need to read into this more in-depth. Appreciate you raising this.
The long-tail point is true. We don't do any tracking in our implementation, but we've been trying hard to refine our total our total permissions approach to think about more edge cases such as this, while not being too annoying. We think this is a general tricky issue with AI alignment ('do what I mean not what I say').