this is my attempt at writing my own coding agent harness.
(1)
the biggest goal with this harness is constant session persistence. every session is backed by a persistence file. this allows you to stop the agent at any point, and pick back up by reopening the agent with the same session file.
usage:
ajail ajent --api-key=<KEY> [your-session-file]
if no session file is provided, one is automatically created in
~/.ajent/sessions/, named after the basename of the current
directory, a timestamp, and a short random hash (e.g.
myproject-20260218-143052-a1b2c3d4.hjl).
to make this work, this harness uses the provider-independent bellman library, which has great support for provider-independent session serialization.
the session files are in the "heredocs json lines" format, which was my idea for the least complicated agent session serialization format that's also kind of human readable. i wrote heredocs json lines documentation.
if you want to switch agent models, you can stop the agent and restart it with the same session file with a new model. you can also make copies and edit and fork, and all of the normal file operations you might do.
i prefer this to a complex ui that has fork operations and so on.
(2)
speaking of complex uis, a secondary goal for this agent is to not be a TUI. we leave the shell alone. if you want command editing history, you can run this with rlwrap.
(3)
like pi, this harness has no concept of privileges or permissions. that is something you should get from the sandbox you run this harness in. i use ajail.
(4)
i've tried to be thoughtful about the ergonomics of the tools for the llms. for instance, reading and editing files uses hashlines.
(5)
there are still many things to do. tool calling output is not very friendly
or formatted, and the web_fetch and web_search tools are hard to do in a
provider independent way, and are currently in need of improvement.
see LICENSE