Vibe Graphing: building editable multi-agent workflows from intent

Watch the source walkthrough on YouTube.

Multi-agent systems often need more than prompts. Someone has to define roles, data passed between steps, branching, tool use, and recovery behavior. That orchestration can become difficult to change once it is spread across implementation code.

MASFactory calls its approach “Vibe Graphing.” It turns a natural-language description into an editable structured representation and then into an executable workflow. The key word is editable: it is a human-in-the-loop process, not a promise that an AI can infer a correct architecture without review.

Three stages, not a black box

The paper describes three stages. Role Assignment identifies the actors and their responsibilities. Structure Design proposes nodes and edges for the workflow. Semantic Completion adds the detailed instructions and relationships needed to run it.

At each stage, a person can inspect, change, and give feedback on the result. MASFactory also provides reuse, context adapters, and a visualizer. This gives teams a structured artifact to discuss instead of requiring every architecture decision to live only in glue code.

What the line-count example actually compares

The paper’s ChatDev example contains several different artifacts. The original ChatDev implementation has 1,511 lines, while the MASFactory reproduction has 1,114. Its interactive Vibe Graphing version has 203 lines. A separate task-specific Vibe Graphing workflow specification has 45 lines.

Those figures show that high-level workflow specifications can be much smaller than an existing orchestration implementation. They do not show that one unchanged ChatDev workflow was mechanically reduced from 1,511 lines to 45, and they do not prove the same reduction for other systems.

Construction cost is not runtime cost

Table 2 measures API cost to construct a workflow, not the cost to execute that workflow in production. In the authors’ setup, Vibe Graphing construction for ChatDev costs $0.26, compared with $3.49 for the low-reasoning Vibe Coding condition and $3.02 for the medium-reasoning condition. The workflow-building model is GPT-5.2; the benchmark workflows use GPT-4o-mini for execution.

For that ChatDev comparison, the lower construction cost is roughly 13.4 times smaller. It does not tell us how much a deployed workflow costs per run, or how a different model, provider, workflow, or review process will behave.

What to take into a real project

An intent-to-workflow system can be useful when it makes routing, dependencies, and responsibilities visible enough for people to review. It should still be tested against the tasks it will run, including error paths, tool failures, data boundaries, and changes to the workflow specification.

MASFactory reports competitive results on seven benchmarks under its stated method. That is evidence for the approach in those evaluations, not a general claim that generated graphs are always cheaper or architecturally superior. The paper also notes that checkpoint and resume support after an interruption is not yet built in, an important limitation for long-running production work.

Treat Vibe Graphing as an interface for drafting and reviewing a system. The generated workflow is a starting point for engineering judgment, tests, and operational controls.

Read the MASFactory paper and official repository for the workflow representations, evaluation setup, and implementation.

Come talk multi-agent systems with me: X, Discord, Telegram.