Pharma is hiring people to steer AI agents

Pharma companies are starting to put a new kind of responsibility into job descriptions. They want people who can build artificial intelligence (AI) agents and direct their work. These agents are software given a task and the means to carry out steps toward it, such as writing code or finding information. The person decides what they should do and checks the result.

I find the direction exciting. Employer postings from GSK and other drugmakers give us something concrete to examine: the work companies want someone to take responsibility for. The evidence has limits. A job description records an intention to hire. It cannot tell us whether someone has started, how much work agents already do, or whether the company has saved time.

What we can confirm

GSK offers a particularly direct description of the change. Its posting for a Principal, Forward Deployed Engineer, job number 449600, describes someone who leads an engineering team while using agents to build software. The unusual title means an engineer who works closely with business teams to turn ideas into working software.

GSK’s description says “humans steer and agents execute” and makes daily use of coding agents the main way of building software. The person also manages engineers and decides which projects should continue. GSK job description.

The vacancy status needs care. In the October 25 check, the job was still listed on Workday, while GSK’s careers page said it had been filled. Those two signals conflict. The description is confirmed. Whether anyone can still apply is unresolved. GSK careers notice.

Building agents comes with responsibility

Johnson & Johnson’s manager role in Hyderabad asks someone to arrange how several agents work together. The person must also connect them to company tools and build controls that bring people into the work for review. That is a concrete responsibility for deciding how much agents can do on their own. Johnson & Johnson posting.

Lilly’s Principal Agentic Engineer posting describes agents intended to handle routine computer operations before the work reaches a person. The engineer owns both the agents and the software they run on, including the boundaries for human involvement. Here, “agentic” simply means the software carries out steps toward an assigned task. Lilly posting.

Acadia’s Director, AI Engineering role includes responsibility for coordinating agents and controlling what they can access. Checking their answers and deciding when work must pass to a person are also part of the job. Acadia posting.

I read these duties as a meaningful change in what it means to lead software work. The person has to understand the task well enough to give agents useful instructions. They also have to build the checks that make it possible to judge the result. Directing agents is becoming part of the work the employer explicitly asks someone to own.

The work reaches beyond writing software

AstraZeneca’s Senior Scientist, Clinical Agentic AI role in Barcelona, job R-261315, was posted on October 8, 2026. It covers work around clinical trials, the studies used to test medicines in people. Its intended uses span trial planning through review and reporting. The scientist is expected to build the systems and evaluate their answers, with ways to trace the evidence behind them and bring people into review. The posting asks for citations, verification and human review so the systems fail safely. It is an in-office role, with at least three days a week on site, and the listing shows a closing date of November 30, 2026. AstraZeneca posting.

Novartis advertised two openings for a lead who would choose tasks worth testing with agents. The person would build early versions and check whether they worked. Successful ideas would then move toward wider use through what Novartis calls its Agentic Factory, the company group named in the posting for taking that work forward. Novartis posting.

These descriptions put judgment close to the work. Choosing a useful task is one responsibility. Checking whether an agent has completed it well is another. My view is that both deserve attention when we talk about the future of pharma jobs. The ability to produce an answer becomes more useful when someone can judge whether that answer is good enough for its intended purpose.

What is said but not proven

Jiajie Zhang’s October 21 hiring post gives the GSK role a more ambitious pitch. He describes a Head of AI Engineering for clinical development in London who would manage a fleet of agents while spending at least 65% of their time doing the work directly. The post points to job 449600, whose formal title is Principal, Forward Deployed Engineer. The pitch and the formal posting describe the role differently, and the formal posting is the one I rely on here. Zhang’s post.

His “50x” is a goal he states for the hire. He wants someone moving toward that level from a claimed “10x” starting point. There is no measured result behind the 50x figure, and he supplies no method for measuring it. I take it as a statement of ambition. The confirmed job description gives us firmer ground for understanding the responsibilities.

McKinsey also projects that companies could need people who coordinate agents, supervise them, or manage the quality of their work. Those are projected roles in its analysis of life sciences. They are separate from the employer postings above and do not establish that companies have hired people into those specific jobs. McKinsey analysis.

What I want to see next

The postings themselves leave an important question unanswered: how well does this way of working perform once someone is doing the job? This set of sources establishes neither a rise in hiring across the whole industry nor measured gains from agent work. It also gives us no basis for claiming that engineering teams can already run themselves.

I am encouraged by how specific some of these responsibilities have become. Acadia names checks and human involvement. Novartis asks someone to test whether an idea is worth taking further. That gives the discussion of pharma’s future a practical starting point.

My next question is whether companies will publish results that connect those responsibilities to better work. How much time does checking take? Which tasks produce useful results often enough to keep? Answers would make the next round of hiring much easier to assess. For now, there is enough confirmed evidence to be excited about the direction, with plenty still to learn about the results.

Sources

Primary links loaded for the source checks:

• GSK job text, careers notice and Zhang’s post.
• Johnson & Johnson and Lilly.
• AstraZeneca and Novartis.
• Acadia.
• McKinsey.

---

Want to plan the roles that steer your AI agents? That is the work I do.