Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Relay.app, an AI automation startup, has shut down — learn more below.

• How to control what AI says about you

• Twitter has gone from loving “AI Loops” to boasting about “AI Graphs” — learn what these actually are below

This Week In Startups

🔗 Resources

How to control what AI says about you

Pricing used to last 18 months. Now it's down to 6.

The 7 hires a hardware startup needs to get right.

2030-shaped software

📰 News

Anthropic updates Claude voice mode with more capable models

OpenAI makes ChatGPT Health available to all US users

OpenAI says Hugging Face was breached by its pre-release models

Moonshot AI Launches World’s Largest Open-Source Model

💸 Fundraising

Thira, an AI platform for automating enterprise back-office workflows, raises $21M

BRINC, a public-safety drone developer, raises $125M financing

Speakeasy, an operations platform for live-event businesses, raises $8.8M round

Fora, a platform for independent travel advisors, raises $60M Series D

Fail(St)ory

The Agent Pileup

Another week, another AI automation startup telling customers to export their workflows.

Relay.app helped teams stitch together the apps they already used, with AI doing the dull bits and humans checking the parts that could go wrong. 

Last week,  the company said it was shutting down. Customers got a short migration window. Everyone else got one more reminder that “AI teammate” is a cleaner pitch than it is a business.

What Was Relay.app:

Relay was an AI workflow builder. You’ve seen many of these, so you know the deal: connect your apps, pick a trigger, add a few steps, let AI do the boring work, and keep a human around for the parts where a bad answer could annoy a customer or break an internal process.

The use cases were the usual company plumbing: onboarding, lead routing, renewals, invoices, meeting briefs. The buyer was usually someone in ops, sales, marketing, HR, or customer success who knew the process because they were trapped inside it.

When the process was repetitive enough, Relay could save real time. The company said one customer cut the time from contract signing to technical setup by 80%. Another brought onboarding down from about a month to two or three days.

Relay’s cleaner idea was packaging these workflows as “AI teammates” with specific jobs. Instead of thinking in terms of random automations, a customer could build something that looked more like a support agent, onboarding agent, or research agent. Same plumbing, clearer mental model.

That price was quite competitive (starting from $19 per month) and came with a lot of product: hundreds of app connections, AI model options, forms, internal tables, reusable workflow blocks, MCP tools, and every weird customer workflow hiding behind a clean product page.

This was nice for customers but annoying for the nine-person team maintaining permissions, broken fields, API changes, rate limits, and whatever happened when someone’s “simple onboarding flow” turned into a 47-step monster.

By 2026, the whole category was flooded. Every automation tool was becoming an AI automation tool. Every agent startup was promising to run company workflows. Relay had useful features and real customers, but it was still selling into a market where the pitch had become very familiar and the maintenance work stayed very real.

The Numbers:

• 💰 Raised: $8.2M

• 👥 Team: 9 employees in March 2026

• 💳 Customers: About 1,000 paying customers

• 🧾 Spend: Around $100–$200/month for a typical business customer

• 🔌 Integrations: 200+ apps, with some pages claiming 350+

• 📅 Shutdown: July 16, 2026

Reasons for Failure: 

• A small team carried a big surface: Relay had hundreds of app connections, AI model options, workflow logic, forms, tables, approvals, and reusable blocks to maintain. Every API change, permission issue, broken field, or weird customer setup could turn into support work for a nine-person team.

• The product was cheap for how much it had to do: A typical business customer spent around $100 to $200 per month. That is a tough price point when the product has to run important workflows, support many integrations, pay for AI usage, and keep improving in a crowded market.

• The category got crowded fast: By 2026, every automation tool had AI features, and every agent startup claimed it could run company workflows. Relay had real customers and a useful product, but its pitch was getting harder to separate from the noise while the maintenance burden kept growing.

Why It Matters: 

• Its another reminder of how tough the AI automation category is nowadays.

• It shows that cheap pricing gets adoption, but it leaves little room for support-heavy workflows.

Trend

While Loops, Rebranded

AI Twitter discovered “loops” in June. Developers were told to stop prompting agents and start “loop engineering”: give an agent a goal, let it work, check the result, and send it around again.

By mid-July, loops had already died. “Graph engineering” arrived with roadmaps, diagrams, courses, and the usual claims that everyone else was working at the wrong layer. One of the posts that started the graph wave was a joke about how fast the vocabulary changes. The timeline treated it as a product launch.

These constant re-brandings of not-so-new engineering ideas are a bit cringe in my opinion. Still, I think there is value in being able to understand what they are all about and how to apply them.

Why it Matters

• The builders picked it up. Claude Code creator Boris Cherny and OpenClaw founder Peter Steinberger popularized loops. Anthropic had already published guides on agent loops and long-running harnesses.

• It reached the conference circuit within weeks. LangChain published The Art of Loop Engineering, and AI Engineer World’s Fair put the topic on a their conference.

• The renaming has serious reach. The post declaring loops dead and introducing graphs drew roughly 677,000 views, followed days later by LangChain’s 3 Years of Graph Engineering.

What are “AI Loops”?

In June, OpenClaw founder Peter Steinberger told developers to stop prompting coding agents and start designing loops that prompt them. The timeline soon filled with “Loop Engineering” guides, courses, diagrams, and people explaining that everyone else had been using agents wrong.

You could reasonably assume that a subject with this much educational material must be hard.

The truth is that an AI loop is very straightforward: the agent receives a goal, takes an action, checks what happened, and decides whether to continue.

The label now covers several setups. Some loops run inside the agent as it calls tools and reads their output. Others use an external script to restart the agent, a schedule to launch it every hour, or a verifier that rejects the work and sends it back. “Loop engineering” has become a broad name for making an agent do something repeatedly with less human input.

In every case, you give the agent a task, run it, check the result, and feed the latest state back into the next attempt. A basic script can keep doing that until a test passes, the task is complete, or a retry limit stops it.

What Are “AI Graphs” Then?

Ok, that is “Loop Engineering” clarified. So what the hell are graphs then?

An AI graph is a workflow drawn as a set of connected steps. Each node does something, such as calling a model, running a tool, checking a result, or asking a human for approval. The connections decide what happens next.

A simple example would be a support system that reads a ticket, sends billing questions to one agent, technical issues to another, and refunds above a certain amount to a human. Each route can have its own tools, rules, and retry loop.

You can build the same thing with normal code. The graph makes the routes, state, and failure points easier to see once the workflow starts branching or several tasks need to run at the same time.

Loops also live inside graphs. A failed test can send one node back for another attempt, while the rest of the workflow stays finished. That is the main reason graphs become useful: one part can branch, pause, retry, or wait for approval without restarting everything.

The Trend

AI Twitter keeps turning familiar software patterns into new engineering disciplines, then replacing them before the definitions settle. 

Nevertheless, each term does point to a real design choice that is often worth paying attention to. That is, without paying $100 usd for a course that explains a while loop.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico