Is adding an agent to a broken process going to cut it?

I've been using AI since 2023, and the evolution since then has been massive, both in what the models can do and in how well they do it.

For me, the turning point came at the end of 2025. Anthropic released Claude Opus 4.5 on 24 November, and by then Claude Code, the coding tool it had first released that February, had matured into a real harness: the layer that gives a model rules, context and limits to work within. A strong model in a mature harness could finally do serious work.

I wasn't the only one who saw it that way. At Rails World 2026 in Austin, David Heinemeier Hansson, who created the Ruby on Rails framework, called this turning point his industry's Kodak Brownie moment. Before the Kodak Brownie appeared in 1900, a realistic painting took one or more expert painters and months, sometimes years, of work. Once a cheap camera could do it, painters had to reinvent their discipline. His point wasn't photography. It was what happens to a profession when a new technology does its core work faster and cheaper, which is the shift developers are going through right now.

Pharma jobs will go through their own version of this. Much of regulatory and quality work today is the hand-painted portrait: compiling, reconciling and cross-checking documents, slowly and by hand. Agents will take on more and more of it. What stays with people is the part that needs understanding: deciding what the work means and whether it's right.

The lesson I took from it is that AI is really good at processing data, calculating and finding patterns, but judgment and understanding are still ours. That's why Roger Penrose likes to call it "artificial cleverness" instead of "artificial intelligence". On its own, it isn't enough: it needs a harness around it, and human judgment to oversee its outputs, authorize actions and make decisions.

In pharma, I've watched good models get it wrong

As the technology improved, I also watched it hallucinate regulation codes and regulatory requirements. I watched it make erroneous extrapolations, carrying a requirement into a context where it doesn't apply. And I watched it confuse the meaning of an acronym.

There are a lot of acronyms in pharma. On top of the industry's own, every company creates its own to refer to day-to-day items. A new hire learns them in the first weeks, by asking. A model can't ask. It only knows what it's given.

Give a model context and the results change

There's a paper that shows this better than any of my anecdotes. In a drug-safety study published in JAMIA Open, researchers gave GPT-4 questions in plain language and asked it to turn them into queries over a structured pharmacovigilance database. With only the database schema, 8.3% of its queries were correct. When they added business context, what the data meant and the rules behind it, 78.3% were correct.

Same model. Same data. The only thing that changed was the context.

Should we add agents, or design systems for them?

So when everyone in pharma and biotech started talking about AI, my question was: is adding an agent to broken processes going to cut it? Or should a system be designed natively for agents?

DHH answered the software version of that question in the same keynote. Apps are rushing to embed chatbots, and he argued they don't need them: they need to be something the user's own agent can work with. In pharma, the bolted-on version looks like this.

An agent added to a process built on documents inherits whatever that process is missing. It reads the same PDFs a person reads, without the years of context that person carries. It doesn't know which version is current, what your acronyms mean, or which requirement belongs to which market.

And when it gets something wrong, whoever signs has to catch it. To sign, you have to judge whether the work is right, and you can only judge what you can trace back to its source.

Pharma needs what software got at the end of 2025: a harness built for the work. A system designed natively for agents starts there. The rules, the context and the limits live in the system, and the data carries its meaning. Every agent works inside the same harness, which decides what it can see, what it can propose and what has to wait for a person. You can change the model next quarter, and the harness stays.

The way we work is changing as we speak. It's being reshaped around how to use this technology, and I think the companies that understand how to get the most out of it will be the ones that succeed. For that, they'll need to rethink how they handle their processes and their information. The sooner they start, the better positioned they'll be.

That answers my question: bolting agents onto broken processes won't cut it. This is why we built Krevyos: an AI workspace where you can bring whichever agents you choose, as first-class citizens collaborating with your people in one pharma-grade environment. The grunt work moves to the agents. The judgment, and the signature, stay with you.

If you're asking yourself the same question, or want to compare notes on where AI really stands today, my messages are open.

Get in touch