Workgraph: A Manual

Workgraph models work as a directed graph: tasks are nodes, dependency edges connect them, and a scheduler moves through the structure by finding what is ready and dispatching agents to do it. Everything — the graph, the agent identities, the configuration — lives in plain files under version control. There is no database. There is no mandatory server.

Dependencies create ordering, parallelism emerges from independence, and structural cycles introduce intentional iteration where work revisits earlier stages. Layered on top of this graph is an agency — a system of composable identities that gives each agent a declared purpose and a set of constraints. Together, the graph and the agency form a coordination system where the work is precisely defined, the workers are explicitly characterized, and improvement is built into the process.

This manual covers the complete system: task graph structure, the agency model, coordination and execution, and the evolution loop that improves agents over time.