A harness runs one agent; a graph decides how many agents, tools, people and facts fit together. This page explains what graph engineering means, the different kinds of graph it covers, how much of Y Combinator's 2026 output does it, and what we expect to come next.
Researched October 2026. This is the companion to our page on AI harnesses and uses the same method. Claims about other people's work cite a source (G1 to G15, listed at the end). The YC figures are our own count from YC's public directory; the method and its limits are written out. The last section is a prediction and is labelled as one.
A graph is nodes joined by edges. In graph engineering the nodes do the work, and a node can be plain code, one model call, a tool call, a human approval, or a whole agent with its own harness. The edges decide what runs next: some are fixed, some are chosen at run time by code or by a model. A shared state travels along the edges, so each node sees what the earlier ones did. G1
The point is control. A single agent decides its own next step, which is right for open-ended work and risky for everything else. A graph lets you say "classify before answering", "look at the repo before proposing a change" or "get approval before anything leaves the building", and have that hold every time. G1 Anthropic drew the same line in 2024: workflows run LLMs and tools "through predefined code paths", while agents direct their own process. Its five workflow patterns (prompt chaining, routing, parallelisation, orchestrator and workers, evaluator and optimiser) are all small graphs. G5
Production graphs are usually not DAGs (directed acyclic graphs, the shape data pipelines use). They have cycles, for retries, validation and waiting on a person. A loop is just the simplest cyclic graph, with one node. G1
One honest complication: "graph" means two different things in AI. Most of the 2026 talk is about control-flow graphs (which step runs next). The older meaning is knowledge graphs (which facts are connected). Both are covered below because products increasingly use both at once.
| Kind | What the graph holds | Examples |
|---|---|---|
| Agent and workflow graphs | Steps and agents as nodes, routing as edges, shared state, checkpoints and human pauses. | LangGraph (65M+ downloads a month, by LangChain's count); Google ADK 2.0, rebuilt in 2026 around a graph-based workflow engine; Microsoft Agent Framework workflows, the successor to AutoGen, whose own GraphFlow is now in maintenance mode G1 G3 G6 G7 |
| Reasoning flows | A fixed sequence of model calls and tests for one kind of problem. | AlphaCodium (Qodo, formerly CodiumAI); Anthropic's workflow patterns G4 G5 |
| Swarms and fan-out | One task split across many agents in parallel, then merged. A graph with a wide middle. | Acyclic Labs, Superset, Emdash, Rindler, Composal and Aster, all YC 2026 (section 4) G14 |
| Knowledge graphs for retrieval (GraphRAG) | Entities and relationships pulled out of documents, so a model can answer questions about a whole collection, not just the closest paragraph. | Microsoft GraphRAG, which beat plain vector search on whole-corpus questions in Microsoft's tests; LazyGraphRAG, which defers the expensive LLM work to query time to cut indexing cost G8 G9 |
| Memory graphs for agents | What an agent has learned, as facts that change over time, with old facts marked out of date rather than deleted. | Zep and its open-source engine Graphiti; YC's Hyper ("a self-maintaining knowledge graph" for AI employees) G10 G14 |
| Workflow DAGs | The older, deterministic lineage: scheduled pipelines whose steps never loop back. | Apache Airflow, where a DAG "encapsulates everything needed to execute a workflow" G11 |
Named LangGraph users on LangChain's own site include Uber, Lyft, Cisco, Nvidia, Coinbase, ServiceNow, Workday, Elastic, Rippling, Harvey and Vanta. G12 That's a vendor's logo wall, so it shows who has adopted it, not how much of their work runs on it.
Graphs lose on open-ended work. GPT Researcher, a popular open-source research agent, replaced its graph-shaped pipeline with a single Deep Agents harness, and LangChain moved its own deep-research agent the same way: planning and delegation were too hard to fix in advance. G1 GraphRAG has its own cost: building the knowledge graph up front is expensive, which is the problem LazyGraphRAG was made to solve. G9 The rule of thumb that falls out: use a graph where the work has a known shape, and put one strong harness inside a node for the parts that don't.
Short answer: very few sell it, and almost none use the word. In our sample about 2% of YC's 2026 companies sell graph or multi-agent orchestration as their product (4% with a looser reading), and another 3% describe a product visibly built as a graph of agents. Nobody publishes an official count, so we made our own, the same way as on the harness page.
We pulled every company in each 2026 batch from YC's public directory (via the open yc-oss mirror,
updated 9 October 2026) and searched each one-liner and description. G14
| Batch | Companies | "Graph" or "ontology" | Multi-agent, swarm or orchestration | "Workflow" |
|---|---|---|---|---|
| Winter 2026 | 198 | 1 | 3 | 22 |
| Spring 2026 | 193 | 2 | 5 | 28 |
| Summer 2026 | 231 | 1 | 9 | 19 |
| Fall 2026 (in progress) | 123 | 0 | 3 | 12 |
| All 2026 | 745 | 4 (0.5%) | 20 (2.7%) | 81 (10.9%) |
The word "graph" barely appears: 4 companies of 745. The middle column over-counts: only 10 of its 20 matches are about coordinating AI agents; the rest are drone swarms, satellites, hardware test rigs, cloud VMs or model routing. "Workflow" is too broad to mean a graph on its own; it mostly means "a business process we automate". For comparison, 36% to 49% of these batches use the word "agent" (harness page).
Keyword counts can't tell what's under the hood, so we re-read the same random sample of 150 companies used on the harness page (drawn with a fixed seed from the 622 companies in Winter, Spring and Summer 2026) and asked a narrower question of each:
| Group | What counts | Companies | Share |
|---|---|---|---|
| Sells graph engineering | A product for building or running graphs of agents or workflows: Jinba (enterprise AI workflows), Pentagon (a workspace where AI employees "coordinate"), Wato (shared memory and reusable workflows across agents) | 3 | 2% |
| Looser reading | Adds tools that fan work out to many parallel agents without saying how they connect (Superset, Emdash) and shared agent memory (Mosaic) | +3 | +2% |
| Built as a graph, visibly | A harness product whose description names several coordinated agents or a fixed multi-stage pipeline: CellType, AutoSitu, TovenAI ("six agents currently live"), Foaster, Cardinal | 5 | 3% |
| No visible graph | Everything else, including most harness products, which may well use graphs inside but don't say so | 139 | 93% |
With a sample this size, the true share selling graph engineering is likely between 0.7% and 5.7% (strict) or 1.8% and 8.5% (loose), and the share with any visible graph between 4% and 13% (95% intervals). The sample had no knowledge graph or GraphRAG company at all; those exist in the batches (section 4) but are rare.
Limits, stated plainly. This measures what companies say, in YC's short marketing descriptions, sorted by one reader (us). It certainly under-counts companies that use graphs: LangGraph alone is downloaded 65 million times a month, and many of the 47% of companies we counted as harness products on the other page probably have a graph inside. G1 What the numbers do show is that, in 2026, graph engineering is something startups use, not something they sell. A second, independent reader would make this stronger.
From YC's 2026 directory, in the companies' own words and with their own figures, which we haven't checked. G14
The named layers so far: prompt, context, harness, loop, graph. G2 Each new name widened the frame, from one call, to one agent, to one agent repeating, to many agents wired together. The same pattern is likely to continue, so the question is what the next, wider frame is.
We think graph engineering is a useful name for something mostly settled, and that the next shift is from drawing graphs to governing them. Over the next 12 to 24 months, with how sure we are:
Send spawns as many workers as a task needs,
and ADK 2.0 offers "dynamic workflows" in plain code. G1 G6 Swarm startups like Acyclic already split
work on the fly. We expect the common case to become an agent that plans a graph, runs it and revises it, with
people reviewing the graph rather than writing it. Graph engineering becomes graph review.What this means for us: use graphs where a process is known (intake, review, approval, payment), keep one strong harness inside the nodes that need judgment, keep our data in open formats a knowledge graph can be built from later, and spend our effort on the checks between nodes.
Provided as is, for information only, with no warranty and no liability. Not legal or professional advice. Security and radio material is for authorized, lawful use only. Full disclaimer.
Visits are logged on our own server: IP address, browser and device details, and pages viewed. We use them to understand traffic and spot attacks, and never share or sell them.