5. Working with Agents

The previous chapter argued that work is moving from how to what. This one is about the practical consequence: you now have a new kind of collaborator on the team — one that is genuinely not a person — and leading it well starts with being clear about which kind of collaboration you are setting up.

Four ways to work with an agent

There are four working models, and the most common mistake is using one where the task called for another.

• Supervised agent. Humans set the goal and check in periodically; the agent executes the detail in between. Genentech's research agent works this way: scientists point it at a target, it proposes hypotheses and does the legwork, and human experts vet what it brings back — and the vetting is only possible because the agent, built on AWS with Anthropic's Claude, returns cited, traceable outputs a scientist can follow back to the source.¹ As Genentech's Aviv Regev frames it, the agent is there to boost the researchers, not stand in for them: it runs the multi-step search and adapts as it learns, while the human keeps the judgment that says which of its answers to trust. Right for well-bounded work where the human's scarce attention is best spent on direction and review.
• Augmented intelligence. The agent sits beside a person as an assistant that amplifies them in the moment — the pair-programming model, where it suggests and the human verifies and refines.² Right for craft work where a skilled human stays firmly in the driver's seat and wants leverage, not autonomy.
• Co-leadership. Humans and agents share the work dynamically, each taking the lead where it is stronger and handing back where it is not. The most demanding model to run, because the handoffs have to be designed — but increasingly where the frontier sits.
• Orchestrated multi-agent. Several specialised agents work together under human coordination. As the researcher Tallam puts it, the real innovation in agentic AI is orchestrated networks of agents designed to slot into human workflows, rather than the lone autonomous agent. Amazon's warehouses are a concrete picture: vision, path-planning, and control agents coordinating behind a single human instruction like "move these to zone B." The instruction is an outcome, not a route — the humans set what they want delivered and the robotics agents work out the how between them, re-optimising the floor continuously as orders shift, without an engineer rewriting the plan each time conditions change.³ That is the declarative pattern at industrial scale: outcome in, self-coordination across the agent network, and human oversight kept for the goal rather than the steps.