Two Pro accounts’ worth of weekly quota, gone — burned through by using Codex to build a 3D website.
Not from writing code. From explaining things over and over.
Every time I switched agents, the new one didn’t know my project. I had to re-tell it: why I made this change, how far along we are, which traps I already fell into. By the time I finished explaining, it would re-read the codebase, and the quota dropped another notch.
A few days ago I used GPT-6 Astra to throw together a 3D personal website. It turned out great — and it drank quota like water. After Codex hit bottom, I stared at the screen and thought: if only the next agent already knew what the previous one did.
Then I stumbled on an open-source project called AgentGit. That’s exactly what it does.
Repo: https://github.com/Einsia/agent-git
Site: agent-git.com
So What Actually Is AgentGit
One line: GitHub, but for agents.
What it records isn’t code commits — it’s an agent’s context, tool calls, and key decisions. These records are called Sessions. They can be saved, handed off, shared, and worked on by multiple people.
Dev work, copywriting, video, research — as long as the task went to an agent, the process can be stored and passed on.
I gave it a spin: let several different agents relay the work and finish that 3D website.
Capability One: Handoff — Relaying Across Agents
Installation was easier than I expected. You just throw AgentGit’s official instructions at an agent and it handles the rest. I gave them to Wesight.
Once installed, I had Wesight analyze what interesting features the 3D site could still get. It came back with a pile of options — day/night and weather cycles, NPC roaming, seasonal and holiday skins.
I saved that conversation into AgentGit, then had Codex read the Session and continue.
That step is Handoff: packaging the requirements, context, and key decisions Wesight organized, and handing them to Codex to implement.
Codex read it and proposed a plan immediately. It looked fine, so I told it to go. Soon the starfield and meteors appeared — the scene shifts with real-world time, cycling through sunrise, sunset, and night sky, and if you’re lucky you catch a meteor shower.
Then Codex’s quota hit bottom too.
I switched to Mimo and had it finish the visitor starfield feature. Mimo read the progress and key decisions Codex left in AgentGit and just kept going.
Switching agents used to feel like onboarding a new hire from scratch. Now it’s more like handing over the desk, the files, and the to-do list — the next person just picks up and continues.
AgentGit currently supports Claude Code, Codex, OpenCode, OpenClaw, Hermes, and WorkBuddy — all of them can do Session handoff and multi-person collaboration.
Capability Two: Remote Control — Multiple People Directing the Same Agent
When I was done, I had the agent save the conversation into a designated repo. Open the repo and you can see all the Sessions stored in the project.
Beyond handing a Session to another agent, you can share it with a friend and set read-only or edit permissions.
I gave a friend edit access. He had his own agent read the context, then added a “seat reservation” feature — now you can sit and rest on the fountain plaza, and if a sprite has taken the seat, it tells you:
Someone’s sitting here. Try another one.
That’s Remote Control: several people can direct the same agent together, and I can watch the edits and results in real time.
Same thing on a team. A colleague hands over a Session, and whoever takes it doesn’t have to chase down half an hour of background first.
Capability Three: Hub — Other People’s Sessions Growing Straight Into Your Project
AgentGit also has a Hub, collecting Sessions people have made public.
Before starting something, you can search whether someone’s already done it, what traps they hit, what prompts they used.
Code can be open-sourced. An agent’s working experience can be open-sourced too — that’s the most interesting part of the Hub.
I found a “footprint star map” case in there, thought it looked fun, and had Codex read and reference that Session to build one for my site. After a bit, it was really there. The star map automatically records every step you take on the planet — and if you’re creative enough, you can even use it to draw.
Click into someone’s Session and you see the whole process: prompts, the agent’s replies, actual operations, key decisions — all of it.
Someone publishes an idea, I have my agent read it, and it grows straight into my project. One person’s experience flows across different agents and different projects just like that.
So I published my footprint star map process too. Next time someone hits a similar need, they can have their agent find my case and skip a few detours.
What It Doesn’t Solve, and Who It’s For
Let’s be honest.
AgentGit is not a Git replacement. It’s a Session-record layer on top of Git — you still commit code the normal way.
It’s also not magic. The Hub’s value depends on how many people are willing to publish their Sessions; in the cold-start phase, what you can find will be limited. And the original write-up doesn’t dig into data storage or privacy policy — if you’re using it on a team, verify that yourself first.
The one thing it genuinely solves: context loss when passing between agents, and between people.
If you only occasionally use an agent to write a small script, you don’t need it. But if you’re like me — quota constantly running dry, bouncing between several agents, or a team taking turns on the same project — the time it saves on repeated explanations is far more than the time it takes to install.
Final Thoughts
Open source used to mean code and final results. In the agent era, the process starts to be worth something.
When I see a stunning case, more than the result, I want to know how the requirements were framed, where the traps were, why it was changed this way, how an idea landed step by step. That experience, buried in chat logs, can finally be saved and reused.
AgentGit is already open source — go take a look on GitHub and drop a Star while you’re there.
How many agents do you bounce between? What’s the most annoying step when you switch? Tell me in the comments.