As AI agents mature, the question becomes: what is the human for?

As AI agents move from demos into real drug-discovery workflows, the question that underpins all others gets harder to avoid: what is the human actually for, once the agents become capable?

Dr Nick Lynch is not a hype merchant on this subject. He spent over 13 years at AstraZeneca, served as a co-founder at the Pistoia Alliance and now runs the data and informatics strategy consultancy Curlew Research. Having witnessed enough technology cycles to be wary of both the doom and the evangelism, he offers a fairly practical model of how mixed teams of people and AI agents might actually work and where the limits sit.

‘AI in the loop’ versus ‘human in the loop’

The phrasing people reach for turns out to matter more than it first appears. The default has been ‘AI in the loop’ but Lynch flags a quiet objection to it.

“Some people have decried that to say it downplays the role of human, while the phrase ’human in the loop’ emphasises that the human is still there… AI is just a component,” he says.

It reads as semantics until you sit with it. ‘Human in the loop’ frames the person as a checkpoint inserted into an automated process; a gate the AI passes through. ‘AI in the loop’ inverts the default: the human workflow is primary, while AI is one contributor among several. For an industry nervous about jobs and autonomy, which metaphor you adopt quietly encodes who is assumed to be in charge. It is not a settled question and Lynch is honest that he is “still struggling a little bit” with it himself.

Agents are not one thing – they are a team with roles

The more concrete idea is that a good agentic workflow is not a single clever model but rather a team, deliberately composed of different functions and responsibilities. Lynch draws the analogy directly to how you build a capable human team:

“You think about all the approaches to building a good team; you need somebody who’s good as a finisher, you need the ideas person…That’s no different to any agentic framework.”

The point is that effective AI workflows shouldn’t take the simple path. They need challenge built in: a broker and a challenger, validators that test rather than rubber-stamp. This is where the conversation sharpens an important distinction. The popular framing is the ‘LLM as judge’ – i.e., one model checking another’s work. Lynch proffers that notion as too weak.

A simple validator is just a filter gate. What richer agentic workflows need is something adversarial.

A validator is a pass/fail gate at the end. An adversarial agent is something else: it argues, it pushes back, it tries to break the proposal. The richer image that emerges from this is not a production line where one agent does the work and another inspects it, but rather agents that gang up, feeding back on one another, tugging each other towards a better answer through genuine disagreement.

A simple validator is just a filter gate. What richer agentic workflows need is something adversarial. 

The flow isn’t one-way: the agents feedback on each other and the collective ends up better than the singletons.

The term, ‘better than the sum of its parts’, is worth acknowledging. The whole concept underpinning multi-agent design is that the interaction between agents produces something none of them would produce alone; the way an hour of human discussion can go in directions no single participant expected.

There is a real drawback to note, however, and it is easy to miss this in the optimism. Adversarial and multi-agent setups are not free; they multiply cost, latency and complexity, plus they can fail in their own ways. Agents can converge on a confident consensus that is “confidently wrong”, or that perform the theatre of disagreement without it changing the outcome. ‘Better than the sum of its parts’ is thus a design goal, not a guarantee; a badly composed team of agents can easily be worse than a single well-prompted model. The human-team analogy cuts both ways, as anyone who has sat in a dysfunctional meeting knows.

The last thread of human value

This is where the conversation gets, in Lynch’s own word, “philosophical” and he is candid that it might not belong in a sober trade piece. But the underlying point is sharp enough to keep.

The thing that still separates a human team from an agentic one, by his account, is the collaborative, inventive spark – the inspiration that comes from people thinking together in the same space.

“We’ve got that intelligence and inspirational and inventive aspects that make teams effective… that’s perhaps the last thread of human value before the AI agents take over,” he said.

The “before the AI agents take over” is delivered lightly, but the anxiety underneath it is real and widely shared. Lynch’s framing is constructive rather than fatalistic: the response to that fear is to make the most of what humans uniquely bring, and use AI where it genuinely helps, rather than pretending the two are interchangeable.

What do humans uniquely bring? The conversation lands on a clean, if deliberately binary, contrast. Large language models are, at root, knowledge and a way of processing it using probabilistic connections. Humans run on something different: instinct, situational understanding and what might fairly be called ‘emotional intelligence’.

At a deliberately binary level: large language models have knowledge and a way of processing it, while humans have instincts and a way of understanding situations.

Neither is clean. The AI’s outputs are probabilistic; the human’s are subjective. Pressed on which failure mode is worse, probabilistic or subjective, the honest answer is that both have issues and that the interesting design space is the combination of both, rather than a contest. We are not at artificial general intelligence (AGI) and the question of whose judgment to trust when remains open.

What to take from this

The temptation with multi-agent AI is to treat ‘a team of agents’ as a magic phrase: more agents, more validation, better answers. The more honest reading from this conversation is that the same issues that plague the effectiveness, or lack thereof, of human teams apply directly: roles have to be deliberately designed, challenge has to be real rather than performed, and the feedback between members is where the value lives – if it lives anywhere.

The unresolved questions are the ones worth carrying rather than resolving prematurely:

• Whether to frame it as ‘human-in-the-loop’ or ‘AI-in-the-loop’
• Whether adversarial agent teams earn their cost
• Precisely what the durable human contribution is once the agents are competent – with instinct and situational judgment being the current best answer, but not a permanent or comfortable one.

Get the composition right and you get the cliché that happens to be true: one plus one equals eleven. Get it wrong and one plus one plus one plus one just equals four, or three; a collection of singletons that cost more than the sum of what they deliver.

Read part I of the conversation to discover where AI is delivering real value in early drug discovery.