June 2, 2026
Introducing Flowtrace: Watch Your AI Agent Work, Step by Step
Today we are open-sourcing Flowtrace — a transparent, reusable way to run a task with your AI agent. It works with any LLM or coding agent, including Claude Code, Codex, and Cursor. It is MIT-licensed, written in Rust, and the same tool we use ourselves at MorphMind.
The problem: the process disappears
Today an AI agent tells you what it plans to do as prose. You read 300 words to find the one wrong premise. To intervene, you compose a paragraph countering it. When the work runs, it is a black box — the agent says “I will do X,” you wait, it returns an answer. You do not know where it is, whether it deviated, or whether it got stuck. And when it is done, only the deliverable remains. The method that produced it — the part an expert carries in their head — is gone.
That is not a capability problem. It is a process problem. And a process you cannot see is one you cannot trust.
A trace makes the work visible — and steerable
A trace turns a kind of task into a graph your agent works through, node by node, while you watch. It is a glance-able shared object: you see the wrong step in 200ms, point at it, and change it. Alignment shifts from natural-language back-and-forth to pointing at a structure both of you can read.
Four properties make a trace different from a script or a prompt:
- Transparent. See the work as a graph and open any step’s output, so you know how a result was reached — not just what it says.
- Grounded. Every step writes a real file, and the agent cannot cite a result that is not on disk. The work is there to open and check, not a claim to take on faith.
- Reusable. The next task of the same kind starts from what you already built — you run the same steps on new inputs instead of starting from zero.
- Evolving. Improve a step once and it carries into every run after, so the trace sharpens as you refine it.
The missing layer of the agent stack
An agent stack has three altitudes, and they compose. Skills are where the moves live — a function, an MCP tool, a playbook. Workflows are where control flow lives — explicit order, branches, gates. Traces are where composition lives — how those moves string together to do a kind of task, as a graph you can see and steer.
That composition knowledge never had a home of its own. Model upgrades lose it. Sessions lose it. Cross-agent transfer loses it. Trace is the layer that holds it — moving us from training agents to teaching them.
How it works
A trace is just a folder backed by git. It declares steps with their dependencies, the assets each step produces, and a final deliverable. Every write the agent makes is one git commit, so the history becomes an audit trail you can time-travel through. Stop at any node, override the agent’s judgment, and every downstream node that depended on it goes stale and re-runs — you intervene at one node instead of redoing the whole task.
Try it
Flowtrace is open source under the MIT license. Clone it, run a reference trace, and watch the graph light up step by step — then tell us where it breaks.