ABOUT WINDAGS

A local-first DAG runtime for multi-agent workflows, built around a 503-skill catalog and a five-stage retrieval cascade for selecting expertise per node.

WHAT IT IS

WinDAGs decomposes a natural-language task into a directed acyclic graph of subtasks, matches each node to a hand-written skill from a 503-skill catalog, and executes the graph in waves of parallelclaude -pprocesses. Each wave's outputs feed forward into the next.

The skill catalog is the differentiator. Vanilla Claude does fine on the first response; the second response forgets the project's conventions. WinDAGs grafts the right specialist into the prompt for each node — a cursor-pagination expert for the GraphQL node, a Postgres failover expert for the database node, and so on — chosen by a five-stage retrieval cascade rather than a vector lookup or a keyword tag.

Source-available under the Business Source License 1.1. github.com/erichowens/workgroup-ai.

INFLUENCES

Each row below is a piece of prior work that maps to a concrete subsystem in the codebase. Where there's a paper or canonical reference, follow the link in the source.

Tradition Influence Map
TraditionContribution to WinDAGs
Graph TheoryDAG structure, dependency resolution, topological sorting
Information RetrievalFive-stage cascade (BM25 + Tool2Vec + RRF + cross-encoder + attribution k-NN) for skill selection
Distributed SystemsWave execution, backpressure, circuit breakers
Constitutional AISelf-governing agent behavior, quality constraints
Software CraftsmanshipFour-layer quality evaluation (Floor/Wall/Ceiling/Envelope)
Operations ResearchCost optimization, resource scheduling, constraint satisfaction
Human-Computer InteractionProgressive revelation, human-in-the-loop checkpoints
Resilience Engineering4D failure classification, graceful degradation
Knowledge ManagementSkill lifecycle, pattern learning, institutional memory
Cooperative MultitaskingWave-based parallel execution: nodes in the same wave run concurrently, downstream waves consume upstream outputs

DESIGN PRINCIPLES

Local-first. The DAG runtime, the skill catalog, the attribution database, and the cascade all live on your machine. No telemetry pipeline (yet — if there ever is one, it'll be opt-in).
One scoring path. The DAG executor, the MCP server, the slash skill, and the shortlister all call the same SkillSearchService.search(). Used to not be true. The bugs were terrible.
Cost-visible before execution. Every DAG returns a per-node estimate to the user before any agent runs. Approve, modify, or abort.
Calibrated copy. Where a thing is shipped, it's named with a code link. Where it's in flight, it's called that. No promises about subsystems that don't yet exist.

BUILT BY CURIOSITECH

WinDAGs is built by Curiositech, Inc., a small team that believes AI orchestration should be transparent, auditable, and cost-conscious.

The framework is released under the Business Source License 1.1 (BSL 1.1). You can read the source, run it, modify it, and use it. The only restriction is offering it as a hosted service.

Request Early Access

JOIN THE BETA

Beta testing Spring 2026. Request early access and help shape the future of multi-agent orchestration.

Request Early Access