durable Graph

A durable Graph is an execution mode you use when you hand an AI agent a task that can't be finished in one shot but takes several steps, and you want it to carry that journey through from beginning to end as a single continuous run. It's a bit like handing a delivery driver a full list of stops and saying, "Make all the rounds, then report back once, at the end." What happens at each stop keeps getting logged along the way, but the person who assigned the job hears nothing until every stop has been made.

While doing this, the agent doesn't touch the original working folder directly — it works on its own separate copy instead. It's like the driver moving goods only within an identical replica of the warehouse, so the original warehouse never gets disturbed. That's why using this feature requires the original working folder to be completely clean, with nothing already messed with, before you start.

What sets it apart from a chatbot conversation that ends with a single short answer is that judgment, execution, and results stay connected without interruption across multiple steps. That's why it's called a "graph" — instead of a single line of conversation, it means a path branching across multiple steps is mapped out in advance, and the run follows that path.

The article explains that "in the terminal TUI, you can turn multi-step agent work (Graph) on and off using the /graph on, /graph off, and /graph <task> commands." It adds that running --graph non-interactively means it "waits until the durable Graph finishes, then prints only the final result." The easy point of confusion: this isn't an interactive feature that shows progress in real time — no response comes until everything is done, and it only works when the target folder is in a clean Git working state.

A durable Graph is an execution mode of a specific tool (Apache Maka), so it can't be reproduced in just any service. Still, to get a feel for the concept:

1. In a terminal, try issuing a command in the form /graph <multi-step task description>. Example: /graph Check this project's test code and fix whatever is failing
2. Before issuing the command, make sure the working folder is clean, with no changes pending.
3. After issuing the command, don't expect progress updates along the way — wait for the final result signaling the task is done.