Build1 publisher3 min readPublished
The money behind low-code agents is better documented than their failure rate
n8n went from $2.5bn to $5.2bn in seven months and SAP is embedding its canvas in Joule Studio, while the single number offered for agent failure, an 88% security incident rate, arrives without a population or a definition.
The Engineer · Build desk
What happened
- Zapier has shipped a product called Agents with more than 7,000 app integrations, and Make, Coze and Dify are all expanding their AI-node feature sets.
- Gartner published its first dedicated Hype Cycle for Agentic AI in April 2026, placing agents at the peak of inflated expectations and calling 2026 the year of disillusionment for the technology.
- The reliability side of the post's case rests on an 88% security incident rate for AI agents, which it attributes to NoCode.Tech's coverage of enterprise incident data.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Once 15 nodes in a 200-node graph are LLM calls, running the workflow and diffing the output no longer returns a clean pass or fail, and no platform in the post has shipped a replacement gate.
- exposure A team citing the 88% incident rate to justify a code-first migration is handing a reviewer a number whose population and incident definition are not published, so nobody in the approval chain can check it.
- decision For an SAP shop, the canvas arrives as part of Joule Studio, so the open question becomes who reviews a graph of LLM calls before it touches live systems.
- contradiction Gartner attributes the coming trough to how agentic AI was sold and bought, which is a weaker foundation for abandoning the canvas than the post's conclusion needs.
An LLM node breaks the cheapest test a workflow platform gives you: run the graph, diff the output against the last run, ship. The example in the dev.to post is a 200-node graph in which every connector is already a black box, and 15 of the nodes are LLM calls that return something different every run [17]. Fifteen in two hundred is 7.5 percent of the nodes [20]. A connector you cannot see into still answers the same way twice. An LLM node does not, and the post's term for the result is model nondeterminism stacked on platform nondeterminism [17].
The comparison the post draws is to the 2010 integration wave: SOA orchestration, ESB tools, Mule, TIBCO, jBPM, Camunda, BPMN modelers, iPaaS, all sold on letting business people wire integrations visually [13]. That ended in abstraction tax, version drift, connector sprawl and the question of who owns the graph when the modeler leaves, and the industry went back to code [14]. One piece survived, the BPMN modeler as a companion to the IDE, and the post's summary is that "modeling survived, implementation did not" [15].
The financial half of the story is the better documented one. n8n at $2.5bn after a $180m Series C in October 2025 and $5.2bn in May 2026 is 2.1 times in about seven months, $2.7bn added [1][2][19]. SAP's stake came with a multi-year deal to embed the workflow canvas inside Joule Studio [3]. n8n's homepage sells the same split the 2010 survivors landed on: "Build visually, go deep with code" [6]. Alongside that, the post counts 195,000-plus GitHub stars and a conference, In The Loop 2026 [4], which measure attention rather than uptime.
The reliability half rests on one figure. The post cites an 88% security incident rate for AI agents and attributes it to NoCode.Tech's coverage of enterprise incident data [10]. It does not give the number of deployments, the time window, or what counts as an incident [21]. Treat it as you would a vendor benchmark table: for 88% to describe your workflows, your agents would need write access to live systems on the same terms as whatever population produced the number, and the post's own gloss is that this is what happens when autonomous decision-making touches live infrastructure without a governance layer [11].
Gartner, as quoted in the post, says the trough "reflects how agentic AI was sold and bought, not whether the technology works" [9]. That is a statement about procurement, and the post uses it to argue that code-first is the winning play [22]. The other supporting evidence is sector-level: Microsoft pausing data-center construction, Anthropic delaying a major release over safety concerns [12]. Neither describes a workflow graph in production.
In my view the 2010 split is still the right default: keep the canvas for modeling and orchestration, and put every nondeterministic step behind a versioned interface you can test with fixtures. A funding round measures what an investor will pay for distribution, and SAP is paying for distribution [2][3]. The post's case for the other side is that a broken BPM graph in 2010 was a maintenance chore, while a broken 400-node agent graph in 2026 is "a silent compliance and security incident" [18].
What to watch
- Whether the enterprise incident data behind the 88% figure is published with a deployment count, a time window and an incident definition.
- Whether n8n or Zapier ships deterministic replay or fixture-based testing for LLM nodes, which would restore the diff test on agent graphs.
- Whether Gartner's next agentic AI hype cycle names workflow platforms specifically instead of agents in general.