The problem it solves
Agent workflows can coordinate tools and produce an answer, but organizations also need to know which evidence was used, which policy applied, who approved the result and what later work may safely reuse. A run log records activity; it does not automatically become trustworthy organizational memory.
Graph Workbench joins those concerns without treating them as the same graph. The execution graph controls the current run. After quality and human gates succeed, a Pack projects selected results into a typed context graph with provenance.
The first-run journey
The bundled Professional Software Delivery Pack demonstrates the full loop without an account, database or model key:
npx graph-workbenchA fresh workspace opens with a zero-key fixture. The workflow runs unit, integration, security and supply-chain checks, pauses for accountable approval, produces an integrity-bound release record and preserves the approved evidence and decisions. A follow-up deployment observation then reads that prior Release object instead of reconstructing it from chat or logs.
What is included
- An executable graph runtime with Agent, function, tool, human, router, join, map, loop, subgraph, wait, escalation and compensation semantics.
- Durable checkpoints, role-owned approvals, retries, recovery paths and integrity-checked audit bundles.
- A context graph for typed objects, relations, versions and production provenance.
- Provider-neutral Agent adapters and a deterministic zero-key runtime.
- Installable Industry Packs containing domain ontology, workflow, roles, tools, quality gates, fixtures, deliverables and context projection.
What it does not replace
Graph Workbench is not a graph database, a general integration catalog or a replacement for domain systems such as source control, an EHR, a broker or robot middleware. It governs the work and the evidence crossing those systems. Production teams connect reviewed adapters while keeping Pack contracts and governance paths stable.
In one sentence: use Graph Workbench when an AI-assisted workflow must produce not only an output, but an explainable, reviewable and reusable organizational result.