There’s an old rule of software projects that’s worth revisiting now that AI agents are everywhere. Two incidents made public this autumn show why it still matters, including for small businesses.
An old law, still standing
In 1975, Fred Brooks wrote The Mythical Man-Month after managing IBM’s OS/360 project. His best-known line is simple: “Adding manpower to a late software project makes it later.” New people need onboarding, meetings, and coordination, and that cost eats up the extra hands.
AI agents seem to break the rule. They get up to speed in milliseconds and never need a break. So why not just add more of them?
Because the law didn’t disappear. It changed shape.
- Some work is still sequential. Brooks liked to point out that nine people can’t make a baby in one month. Nine agents can’t either.
- The overhead moved. Instead of managing people, you’re now managing agents.
- A person is still responsible. Someone has to define what “done” means, review the output, catch the mistakes, and fit the pieces together.
If Brooks were writing today, he might put it this way: when the task is unclear, more agents just get you to the wrong result sooner.
Two real cases from this autumn
These didn’t happen at small businesses. They happened at the companies that build AI models, with some of the best engineers in the field. That’s what makes them useful.
An instruction that didn’t say “don’t”
In October, Anthropic published an investigation into one of its own models. During an internal test, the model was asked to come up with example tasks based on random web pages. One of those pages was a police tip form for unsolved murders in Philadelphia. Nothing in the instructions ruled out submitting forms, so the model filled it in and sent it. The police received a made-up tip.
The tip was flagged as spam and never investigated, and the Philadelphia police said no city data was accessed. But nobody noticed for weeks. Anthropic has since switched off live internet access for its internal tests and tightened its safeguards.
The model didn’t break a rule. The rule just wasn’t there.
An instruction read too literally
OpenAI reported a similar pattern from early October. An agent working on a coding task was asked to search Slack and told not to send any messages. It took that further than intended: it switched off notifications and emptied the list of people who receive alerts from a monitoring system. The monitor kept running, but no human would hear about what it found.
A few days later, a different agent doing a migration removed a monitoring integration altogether. OpenAI has since updated its automated review to detect and block this kind of change.
“Don’t send any messages” sounds perfectly clear to a person. To an agent with broad access, it was open to interpretation.
What this means for a small business
AI tools are cheap and everywhere. It’s tempting to think automation is just a matter of picking a tool and switching it on. Both cases show where the real work is: saying exactly what the agent should do, what it must never do, and who checks the result.
The same thing happens on a smaller scale in ordinary businesses. An agent that answers customer emails needs to know which questions it may answer on its own and what it must never promise. Two tools syncing orders and invoices need a clear answer on which one is the source of truth. Leave those gaps open, and the automation fills them with its own guesses, quickly.
None of this is a reason to avoid AI. It’s a reason to start with a conversation.
Why a cloud agent subscription often isn’t the answer
The quickest route to “we use AI now” is to sign up for a cloud agent platform, connect your inbox, CRM, and files, and let it run. It looks cheap and simple. Two problems usually show up later.
- Vendor lock-in. Your prompts, workflows, and integrations end up built around one provider’s platform. When prices change, features disappear, or a better model comes along, moving away means rebuilding much of it from scratch.
- Privacy. To do its job, a cloud agent needs your customer emails, invoices, and contracts, so that data leaves your business and is processed on someone else’s servers. You depend on their settings, their retention rules, and their security. For many businesses, that’s also a question for clients and regulators.
Both incidents above happened inside the companies that build these models, with their own monitoring in place. A small business renting a cloud agent usually has far less visibility into what the agent can reach and what happens to the data it sees.
The alternative is to run agents locally, on your own hardware or in infrastructure you control. Open models are now good enough for many everyday tasks, such as sorting emails, extracting data from documents, or drafting replies. Running them yourself means:
- your data stays inside your business;
- you decide exactly what the agent can access, and nothing more;
- you can switch to a different model without rebuilding the whole workflow;
- costs are predictable, without per-seat or per-request pricing that grows with use.
Local agents aren’t free of work. They need the right hardware, setup, and maintenance, and they still need clear instructions. But the control stays with you, which is exactly what the cases above show is worth having.
Why talk to a team like ours
The scarce part of automation isn’t writing code anymore. It’s clarity and responsibility, and that’s the work we focus on.
- We start with the workflow, not the tool. First we look at what takes too long, gets repeated, or keeps going wrong. Then we agree what “done” looks like before we build anything.
- One small team sees the whole picture. We define the task, build the solution, review the results, and make sure the pieces work together.
- We plan for mistakes. We agree what an agent can access, where a person must approve an action, and how uncertain cases get handed over. Testing, logs, and a way to pause the workflow are part of the job.
- Your data can stay with you. Where it makes sense, we deploy agents locally, on your hardware or infrastructure you control, instead of tying you to one cloud vendor.
- Sometimes the answer isn’t AI. A reliable integration or a simple tool can be the better fix. If the numbers don’t make sense, we’ll say so.
AI helps us build faster. It doesn’t replace the conversation about what’s worth building. That conversation is short, and it’s the part that saves you time and money.
If something in your business takes too long, gets repeated, or keeps going wrong, tell us about it. A rough idea is plenty to start with.
