VineeTagarwaL-code/graph-engineering
7 stars · Last commit 2026-07-23
Claude agent skill that plans multi-step work as a dependency graph. Subagent fan-out, layered fan-in, model tiering, verification retry loops, human approval gates. No framework, just instructions.
README preview
<p align="center">
<img src="assets/banner.png" alt="Graph Engineering" width="100%" />
</p>
**A skill that plans multi-step work as a dependency graph instead of a linear chain.**
No framework. No runner. No dependencies. Just instructions that change how the model plans: which work fans out to parallel subagents, which must wait, where results consolidate before they overflow, and which single node stops for a human.
[](LICENSE)
---
## The problem
Give an agent a big task ("audit these 60 files", "research these 30 companies") and it narrates a chain: first this, then this, then this. Two things go wrong.
1. **Context burns linearly.** By item 40, the context window is full of items 1 through 39. The first fifteen items get careful treatment, the last twenty get a sentence each. You won't notice, because the output still *looks* complete.
2. **Independent work runs sequentially.** Most "and then"s in a plan are narration, not dependency. The steps could have fanned out. Instead they queue.
The fix isn't a bigger context window. It's structure: a dependency graph, executed in phases, with consolidation layers that don't lose detail and gates that catch problems before they ship.