I have Paperclip running and one agent set up. I want a place to give agents ongoing responsibilities, hand them work, and come back to something I can read and use.
The rest of the crew still needs onboarding. I need to give the researcher, writer, and reviewer instructions that make it clear what each should produce and how the work gets passed along.
Connecting the Agent
I started with Codex. One thing that caught my attention was Paperclip’s support for ACP, the Agent Client Protocol. It standardizes communication between coding agents and their clients. Paperclip uses it as one way to run an agent and receive structured updates about its work.
The interface assigning the work can be separate from the agent doing it. Paperclip’s adapter documentation describes both ACP execution and CLI integrations, with different levels of detail available while a run is in progress.
There are a few concerns. The model picker showed GPT-5.6 Sol, but I didn’t see GPT-6 Sol. I’ll use the available option for now. I haven’t worked out whether that’s a limitation of the installed version, the adapter, or how the model list is populated.
The sandbox setting bothered me more. My initial setup worked with the bypass option enabled. I tried unchecking it, and the agent didn’t work. I haven’t diagnosed the failure yet, but hopefully Paperclip does not require bypassing the sandbox in every configuration.
I also haven’t tried changing the execution setting from ACP to Codex CLI. That seems worth testing before deciding how I want the rest of the agents configured. I’d like to understand which part failed and whether I can keep the sandbox enabled.
The Files Matter Too
Getting the NFS storage mounted was useful. NFS lets a machine access files on network storage through a mounted directory. The setup now gives Hermes and Paperclip their own folders there, so work saved in those folders can live separately from the machine running the agents.
I can open the share over SMB when I want to review the files. I also like that Paperclip lets me download the Markdown files it creates. Sometimes I want to browse a folder; sometimes I just want the document from the task I’m looking at.
For all my work, Markdown is already the format I use. A draft can come back into my existing writing process, where I can edit it, check its sources, and decide whether it’s ready. Having another interface for managing agents is more appealing when the resulting work is easy to take with me.
Some of the recent wiki work has been about juggling Markdown files, so it’s good to see that Paperclip has put some work into solving these problems.
Onboarding the Rest of the Crew
Next I want to try the whole workflow: I pass in a brief, the researcher gathers sources, the writer creates a draft, and the reviewer checks it. The reviewer will need to know what I care about before approving anything.
I can picture building my first little factory with it. Keeping the workflow outside a single harness seems more flexible to me, especially if I want different stages to use different tools.
The next experiment is whether those agents can hand work to each other and produce a draft I want to use.
I’d appreciate a follow. You can subscribe with your email below. The emails go out once a week, or you can find me on Mastodon at @[email protected].