Giving AI agents a process the whole team can see
For a research consultancy we built two AI agents, a project manager and an epidemiologist. The most valuable idea was making their way of working visible and editable.
A research consultancy has two kinds of work that deserve equal care. There is the science: study design, protocols, analysis plans. And there is the running of each engagement: plans, timelines, records of decisions, and the many small things that keep a project on course.
We built Apollo for a life-sciences consultancy: a workspace with an AI agent for each. The project manager keeps engagements organised. The epidemiologist supports study methods. Both prepare work for the scientists who make the decisions. This article shares the ideas that shaped them, which we think are useful to any expert team considering AI.
A process people can read
The usual way to tell an AI agent how to work is to write instructions that sit out of sight. We wanted something more open.
Each agent follows a process laid out as a series of connected steps, which the team can open, read and edit. When the firm refines how it works, staff update the process themselves, and the agents follow. Every change is kept in a version history.
This has a lovely effect. The people who know the work best can see exactly what the agents have been asked to do, and can improve it. The agents become a reflection of the team’s own standards.
When we improve the standard process ourselves, teams that have not customised theirs receive the update. Anything a team has taught its agents is always preserved.
The project manager: keeping engagements organised
The project manager sets up each new project in the firm’s own way of working: the project board, the timeline, the outline for a kick-off meeting. It drafts meeting notes and keeps a record of decisions, so that important choices are written down where everyone can find them.
The project manager also keeps an eye on progress. If a milestone is slipping, someone has more on their plate than they can manage, a task is blocked, or a client has been waiting for a reply, it notices.
When it does, it brings the matter to the right person in a clear and consistent way:
- What has happened.
- Why it matters.
- What it has already done.
- The decision it needs.
That structure turns an alert into a helpful request. The person receiving it can decide quickly, with everything they need in front of them.
The epidemiologist: methods drafted and reviewed
The epidemiologist prepares methods work on the firm’s templates: study specifications, causal diagrams, sections of a protocol, summaries of the literature.
Two habits make its work easy to rely on.
It is honest about gaps. A study specification has several required components. If the brief does not give enough to complete one, the agent marks it as open and says what is missing. A clearly marked gap is far more useful to a scientist than a confident guess.
Every draft is reviewed separately. A distinct review step examines each draft against a checklist of the common sources of bias in observational research, and against a recognised reporting standard. The result is one of three: pass, pass with points to consider, or back for another round. The points raised stay on the record until they are resolved.
The aim, in the words of the agent’s own guidance, is work that someone can verify in thirty seconds.
Working together smoothly
When one agent needs something from the other, the request is tracked until it is answered. If a reply does not come, the request is sent again, and after that it is raised on the project where a person will see it. Nothing is quietly forgotten.
Each agent has its own memory, and both share the team’s wiki. The agents remember things deliberately. We chose this over automatic note-taking because a small number of well-chosen notes is more useful than a large number of incidental ones.
Giving each task the right context
An agent works best when it is given the instructions relevant to the task at hand. An agent answering a quick question does not need its full guidance for setting up a new project.
So each kind of task brings only the steps that apply to it. This keeps the agents focused and responsive, and it is a useful principle more generally: clarity often comes from deciding what to leave out.
Clear boundaries
The agents can read the firm’s project boards and conversations so that they stay informed. They cannot change them. That limit is part of the software itself.
Conversations that include people outside the firm are left out entirely. The epidemiologist works only with aggregated, de-identified or synthetic data, and the interpretation of real results is reserved for a named scientist.
And only people speak to the firm’s clients. The agents draft, check and bring things forward. Anything that leaves the firm is reviewed and sent by a person.
Assistants to the experts
Both agents finish their work in the same place: ready for a person to take forward.
We think this is exactly what an expert team should look for. Agents whose work can be checked quickly, whose way of working is open for the team to read and shape, and who understand clearly where their role ends. With that in place, they become a real support to the people whose judgment the work depends on.