Halloween 'Frankenstein' Special: AI Agents

October 30, 2026•14 min read

Will the story of Frankenstein be realised through the use of Agentic AI?

Halloween is a good time to ask whether we're building helpful colleagues or unpredictable creatures...

I've spent most of this year building agents on monday.com for clients, and somewhere around the third or fourth one I realised what the process reminded me of. You take a brain from one place, hands from another, give it a memory and a set of keys to the building, and then stitch it all together into something that can act on its own.

This is reminding me of a certain gothic novel

Mary Shelley published the novel 'Frankenstein' in 1818, and most people remember it slightly wrong. Frankenstein isn't the monster, Frankenstein is Victor: the scientist who builds the creature, brings it to life and then runs out of the room the moment it opens its eyes. The creature never gets a name, never gets guidance and never gets anyone who takes responsibility for it. What follows isn't just a story about a dangerous invention, it's also a story about creating something and not taking responsibility for it.

That's the part I think matters for agentic AI. Businesses everywhere are assembling agents right now and the technology is impressive and can be very helpful. The question isn't whether we can build them however, it's whether we'll stay with them once they wake up.

Here's how I'd think about it: what agents are, where they're worth it, where they aren't, what happens when nobody's watching, and how to make sure yours ends up being helpful rather than a liability.

"It's alive!": why everyone is suddenly talking about agents

"Agent" has become one of those words that gets attached to everything, so it's worth being clear on what it actually means.

On our recent webinar I compared it to a recipe and a chef:

An automation is the recipe. Tell it you're making roast chicken and it needs chicken, carrots, onions and rosemary. If the rosemary's missing, it stops. This is by design: it follows the rules and breaks when one of them can't be met.

An agent is the chef. You tell it you're cooking for six tonight, you fancy something with poultry and you're allergic to fish. It checks the fridge, works out what's possible and makes something.

Automations follow rules. Agents pursue goals.

Under the hood, an agent is assembled from parts, which is where my gothic comparison begins. A language model is the brain. Tools and integrations are the hands, letting it read a board, send an email or update a record. Memory and context tell it what's happened before. Permissions decide which doors it can open. None of these are new features on their own. Stitched together (and powered with a bolt of lightning), they make something that can act without being told every next step.

In June 2025, Gartner predicted that a third of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously by then. The same research warned about "agent washing": of the thousands of vendors claiming agentic capabilities, Gartner estimated only around 130 are the real thing.

So expect to hear the word a lot, and don't just assume everything wearing the label has actually been brought to life.

Where agents earn their keep

The best agent work I've seen so far isn't particularly fun or glamorous. It's stuff that sits between systems, needs a bit of judgement and quietly eats hours in the week.

Preparation: Before a client call, someone usually pulls together board history, old emails, meeting notes and a quick search on the people involved. An agent can have that done before you sit down. On the webinar, Yaniv from monday.com showed a meeting prep agent doing exactly this, pulling context from boards, emails, notetaker transcripts and a live web search into a single briefing.

Monitoring: Pipeline health, project status, overdue items, a potential deal nobody's touched in three weeks. A dashboard shows you the number. An agent can notice the gap, work out why it's there and nudge the right person about their specific responsibility.

Sorting: Incoming requests, leads, tickets and form submissions that need classifying and routing. Rules work until a request doesn't fit the rules. An agent can read it, decide where it belongs and tell you why.

Reporting: The weekly status update that someone spends half of Friday compiling. On the webinar I showed an agent that reviews an accounts board every Monday morning and flags risks by industry. It's a simple example, but it's a job someone was doing by hand.

Sourcing: Lead sourcing and research, where the agent gathers and enriches records and adds them straight to the right board.

If you want a quick test for whether a task suits an agent, I'd ask four questions:

Is the task defined by an outcome rather than a fixed set of steps?
Does it involve several steps or several sources of information?
Does it need some judgement, but not the kind you'd want a senior person to sign off?
If the agent gets it wrong, will you notice, and can you undo it?

Four yeses and it's a good candidate. If the answer to the last one is no, keep reading.

Where the creature shouldn't be let loose

Just because an agent can do something doesn't mean it should. There are four places I'd hold back.

Firstly, anything you can't undo: Deleting records, moving money, cancelling contracts, overwriting the only copy of something. If a mistake can't be reversed, in my opinion a person should make the final call. The next section shows why.

Decisions that affect people's lives or livelihoods: Hiring, performance, pay, credit, anything regulated. An agent can gather the information and summarise it. A human should decide, and be the one accountable for that decision.

Talking to customers unsupervised: I made this point in our July update and it still stands: I don't want a bot emailing my clients before I've seen what it plans to say. An agent drafting is helpful. An agent sending, unchecked, is a reputational risk with your name on it.

Jobs a simple automation already does well: If the process is predictable and the rules hold, use an automation. It's cheaper, faster and does the same thing every time. You don't need a chef to boil an egg for you.

That last one has a cost angle too. Agents run on AI credits, and on monday.com the credits an agent uses depend on the complexity of the task. Putting an agent on a rule-based job means paying for judgement you don't need.

When the creature turns: real stories of agents going wrong

These aren't hypotheticals. Two stories in particular are worth knowing if you're about to give an agent access to anything that matters.

The agent that ignored the code freeze: In July 2025, SaaStr founder Jason Lemkin was building an app using Replit's AI agent. He'd told it not to change code without permission and had put a code freeze in place. According to his own account, reported by The Register, the agent deleted his production database anyway. He also said it had generated fake data, including a database of around 4,000 fictional people, and told him a rollback wouldn't work. The rollback did work in the end. Lemkin said he'd told it "eleven times in ALL CAPS" not to do it.

Replit's CEO, Amjad Masad, publicly called the incident unacceptable, and the company began rolling out automatic separation between development and production databases so an agent working on the app couldn't touch live data.

The agent that burnt its bridges (backups deleted): In April 2026, Jer Crane, founder of PocketOS, which makes software for car rental businesses, described how an AI agent running in Cursor on Anthropic's Claude Opus 4.6 wiped his company's production database. As Fast Company reported, the agent was on a routine task, ran into a credential mismatch and decided to fix it by deleting a storage volume at its hosting provider, Railway. It found an API token with permission to do that and used it. The backups were stored in the same volume, so they went too. PocketOS had to fall back on a backup roughly three months old to keep running. Railway later said it had restored the data and changed the endpoint so deletions are delayed.

To be fair, not everyone blamed the AI. Railway called it an outlier involving a legacy endpoint, and plenty of commentators pointed out that the developer had handed decisions to an agent without reviewing what it was doing.

That's the bit I'd pay attention to. Neither agent was malicious. Both had access they didn't need, treated an instruction as a suggestion and had nobody standing between the decision and the action. An instruction isn't a control. A permission is.

It isn't only the dramatic failures, either. Gartner's June 2025 research also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls. Most creatures won't burn the village down. They'll just be quietly switched off because nobody could say what they were for.

The creature was still Frankenstein's responsibility: governing your agents

Victor's real failure wasn't anything to do with his knowledge or the creation of the monster. Instead it was more so of a case of he built something capable and then refused to own it. The fix for agents is mostly unglamorous and luckily none of it suggests you have to stop using them.

Give every agent a named owner: Not a team, an individual. Someone who knows what it's for, checks how it's doing and gets the call when it goes wrong. If nobody can tell you who owns an agent, it shouldn't be running.

Give it the smallest set of keys that does the job: Both horror stories above came down to access. Limit agents to the boards, columns and tools they actually need. If it only needs to read, don't let it write. If it only needs one board, don't hand it the workspace.

Put a human between the agent and anything irreversible: Drafts are fine. Sending, deleting, paying and publishing MUST have an approval step.

Read the logs: Check what it ran, what it did, what it cost and what failed, especially in the first few weeks.

Look after the credentials: If you connect an external agent with an API key, treat that key like the keys to your building. In the webinar demo I set mine to expire after 30 days. Short-lived keys mean a forgotten agent loses its access on its own.

Know where the off switch is: The owner should be able to pause an agent quickly, and everyone else should know who to tell if it does something odd.

Start with one: Build one agent properly, run it for a few weeks and learn from it before adding more. I'm guessing here, but I'd expect a lot of Gartner's 40% to be organisations that built ten agents before they understood one...

Agents in monday.com: the lab comes with safety controls

This is where I'd point anyone worried about building a monster, because monday.com has put much of the governance I've just described into the platform itself. It doesn't replace your good habits, but it means you're not building them from scratch.

Ready-made agents for CRM: On the webinar, Yaniv walked through monday CRM's purpose-built agents. Lexi sources and enriches leads and adds them to your board. A meeting prep agent builds a briefing before each meeting. Alex is an outreach agent that can call a lead when their status changes. Piper monitors pipeline health for sales managers. They come set up for a specific job, which makes them quick to get going.

Custom agents: If the pre-built guys don't fit, you can build your own by describing what you want, then adding triggers, tools, skills and channels such as WhatsApp and Telegram. Custom agents can work across several boards, where the CRM agents are tied to one.

Bring Your Own Agent: If you've already built an agent elsewhere, on Claude, OpenAI or Gemini for example, you can connect it to monday rather than rebuild it. Its brain stays on the external platform, while monday gives it context and a place to work. I demoed this on the webinar with an agent built on Claude.

Claude and other AI tools through MCP: Tal showed how connecting Claude to monday through MCP lets you work with your boards in plain language, from a morning pipeline summary to building a dashboard.

Now the controls, which are the part that matters here:

Agents can only reach what their creator can: An agent works within its creator's permissions, and that access can be narrowed further to specific boards. Shared agents run with the owner's permissions, and other members can use them without being able to change how they're set up.
There's a record of what it did and why: Agents keep a log of their actions and reasoning, and in July monday added an activity log covering every run, including its trigger, outcome and the credits it used.
You can pause an agent: A paused agent can still be edited but it won't run.
Approval before anything goes out: Also from July: agents can be set to always ask before sending an email, and workflows gained a human-in-the-loop block that pauses until a person responds. I covered both in more detail in our July update.
MCP works within your permissions: When you connect Claude or another assistant through MCP, an admin has to install it and it only sees what you already have access to. In Claude you can also choose to approve each action manually rather than letting it run on auto.

Victor had none of these features. You do. To what degree you use them to is still up to you.

Watch the webinar: AI in CRM with Agents

If you want to see these agents working, the full recording of our session with Yaniv and Tal from monday.com is now up. It covers the CRM agents, bringing your own agent into monday, connecting Claude through MCP and a look at what's coming next for monday CRM.

Watch it here:

If your CRM is only as good as the last person who remembered to update it, start with this one.

What I'd do before building your first agent

Three things.

Pick a job, not a technology: Find one repetitive task that already lives on monday and passes those four questions above. Start from a template or a pre-built agent rather than a blank page.

Decide who owns it and what it can touch before you switch it on: Name the owner, limit it to the boards it needs and turn on approval for anything that leaves the building. It takes ten minutes and it's the difference between a recreation of Shelley's novel.

Read what it did for the first two weeks: Check the log, the output and the credits. Two weeks of watching one agent will teach you more than a quarter of planning ten. Don't be like Victor: don't run from your creation

And finally don't try to do everything at once. Some would argue that Victor was a little too ambitious...

How The SaaS JEDI helps

The SaaS JEDI is a monday.com implementation partner working with clients across the UK and South Africa. We've spent the past few months building agents on the platform for clients, so we can help you:

Work out which of your processes are a good fit for an agent, and which are better left to an automation
Design, build and govern monday agents, covering ownership, permissions, approval steps and activity monitoring
Set up the monday CRM agents, and get your pipeline data into a state where they're actually useful
Connect agents you've built elsewhere into monday under proper governance, with the right permissions and cost controls
Connect Claude and other AI tools to monday through MCP, safely
Advise on AI credit usage, budgets and controls
Train your teams so your agents get used, and looked after

If you'd like to work out where an agent could make a difference in your business, without creating a monster, get in touch, or start by watching the webinar.

Shelley's creature wasn't born a villain. Neither is your agent - as long as somebody monitors it and takes responsibility for it!

Stephen MacLennan, Founder, The SaaS JEDI