Teach Claude to design multi-agent workflows as dependency graphs, not linear pipelines: fan-out what's independent, gate what's untrusted, and spend expensive tokens only where judgment lives.
Reference: "Graph Engineering with Claude: 14-Step Roadmap" by @0xCodez.
Most agent workflows are accidental pipelines: step B runs after step A only because you wrote it after step A, not because B consumes A's output. Graph engineering makes the dependency structure explicit:
- Node = one agent, one bounded job, one input in, one output out.
- Edge = a real data dependency — nothing else.
- Everything without an edge between it runs in parallel.
- Orchestration, dedup, routing, and transforms are code, not agents — edges are free; tokens are not.
The workhorse shape is the diamond:
graph LR
A[Split] --> B[Worker 1]
A --> C[Worker 2]
A --> D[Worker 3]
B --> E[Verify gate]
C --> E
D --> E
E --> F[Merge / synthesize]
- Fan-out / fan-in, and when a barrier is actually justified
- The diamond topology
- Conditional routing in plain code
- Verification gates: adversarial refuters, perspective-diverse lenses, judge panels
- Per-node failure isolation
- Convergent loop-until-dry cycles (dedup against all seen, not just confirmed)
- Pipeline-over-barriers streaming
- "Edges are free" — no-agent-needed as a design win
- Model tiering with a concrete allocation table:
| Node type | Model | Effort |
|---|---|---|
| Orchestration (routing, dispatch, merge) | Most capable available | low |
| Planning / investigation | Most capable available | xhigh |
| Implementation / execution | Second tier | medium |
Plus a pre-ship checklist for any workflow script.
The skill also self-triggers when your prompt sequences steps with "and then" / "next" — Claude checks whether the steps truly depend on each other before defaulting to a linear pipeline.
As a plugin (recommended):
/plugin marketplace add HM-Li/graph-engineering
/plugin install graph-engineering@graph-engineering
Manual: copy skills/graph-engineering/ into your project's .claude/skills/
(or ~/.claude/skills/ for all projects):
mkdir -p ~/.claude/skills
cp -r skills/graph-engineering ~/.claude/skills/Key reference: @0xCodez's Graph Engineering roadmap. This repo packages the ideas as an installable Claude Code skill and adds the per-node-type model/effort allocation policy.
MIT