Mesh Flow is a skill that moves multi-agent orchestration logic out of prompts and into a structured, artifact-driven DAG (directed acyclic graph) framework. Instead of relying on implicit prompt-chaining, developers define their workflow in a single project.yaml file, identifying each node by the artifacts it consumes and produces. This topology-first design decouples workflow structure from the logic inside individual skill nodes.

Before any execution begins, a compiler validates the workflow for circular dependencies, missing artifact producers, and schema violations. This compile-then-run approach catches structural errors early rather than at runtime. Once validated, execution proceeds through a state machine that distinguishes three explicit failure states — failed, blocked, and rejected — each with its own automated recovery path.

Mesh Flow enforces hard gates that models cannot bypass. These include human-in-the-loop approval steps, artifact presence checks, and custom verifier functions. All skill nodes share a standardized adapter contract, which guarantees consistent trace logging and error reporting across the entire workflow.

A CLI is included for local validation, workflow execution, and generating Mermaid-format visualizations of the agent’s decision graph. These visualizations make it straightforward to inspect and communicate workflow topology before deployment.

Mesh Flow is suited to production-grade agentic systems where workflow predictability, auditability, and controlled failure handling matter. It is not a simple prompt-chaining helper and carries meaningful structural overhead for small or single-step agent tasks.

Use cases

Define a multi-step agent pipeline in project.yaml with explicit artifact dependencies between nodes

Validate a complex agent workflow for circular dependencies and missing producers before running it

Insert human-in-the-loop approval gates that cannot be skipped by model outputs

Trace and audit every node's inputs, outputs, and failure state across a multi-agent run

Generate Mermaid visualizations of an agent's DAG to inspect or document workflow topology

Implement custom verifier functions to enforce business rules at specific workflow checkpoints

When to use it

Building production multi-agent workflows where reproducibility and auditability are required

Workflows that need hard gates such as human approvals or artifact presence checks that models must not bypass

Teams that need to inspect and validate workflow structure before any execution occurs

Agentic systems requiring clear, distinguishable failure modes and automated recovery paths

When not to use it

Simple single-step or two-step agent tasks where DAG overhead is not justified

Projects that require a lightweight prompt-chaining library with no compile step

Environments without CLI access, as the included CLI is part of the core workflow

Use cases that do not involve artifact-passing between nodes, since the framework centers on artifact contracts