Giving AI agents a process the whole team can see

We built Apollo for a life-sciences research consultancy: a workspace with two AI agents. The project manager runs the operational side of each engagement. The epidemiologist drafts methods work. Both prepare work for the scientists who make the decisions, and neither speaks to a client. The idea that shaped both is that the agents’ way of working is a document the team can open, read and change.

A research consultancy runs two kinds of work side by side. There is the science: study design, protocols, analysis plans. And there is the running of each engagement: plans, timelines, decision records, and the small things that keep a project on course. Apollo has an agent for each.

Can a team see and edit how its AI agents work?

In Apollo, yes. Instructions for an AI agent normally sit out of sight in a prompt. Apollo’s agents follow a process laid out as a series of connected steps that the team can open, read and edit. When the firm changes how it works, staff update the process themselves and the agents follow it. Every change is kept in a version history.

The people who know the work best can see exactly what the agents have been told to do, and change it. The agents end up holding the team’s own standards.

When we improve the standard process, teams that have not customised theirs receive the update. Anything a team has taught its agents is preserved.

What does an AI project manager do in a research consultancy?

It sets up each project, keeps the record, and raises problems with the right person. For each new project, it builds the firm’s standard structure: the project board, the timeline, the outline for a kick-off meeting. It drafts meeting notes and keeps a decision record, so important choices are written down where everyone can find them.

It also watches progress. It flags four things: a milestone slipping, someone with more on their plate than they can manage, a blocked task, and a client waiting on a reply.

Every flag reaches the right person in the same four parts:

1. What has happened.
2. Why it matters.
3. What it has already done.
4. The decision it needs.

That turns an alert into a request someone can answer in one pass, with everything they need in front of them.

What does the AI epidemiologist draft, and who checks it?

It drafts methods work on the firm’s templates, and a separate review step checks every draft. The work covers study specifications, causal diagrams, sections of a protocol and literature summaries. Two rules govern it.

It marks gaps instead of filling them. A study specification has several required components. If the brief does not give enough to complete one, the agent marks it open and says what is missing. A marked gap is more use to a scientist than a confident guess.

Every draft is reviewed separately. A distinct review step checks each draft against a checklist of the common sources of bias in observational research, and against a recognised reporting standard. It returns one of three results: pass, pass with points to consider, or back for another round. The points raised stay on the record until they are resolved.

The standard, in the words of the agent’s own guidance: work someone can verify in thirty seconds.

How do the two agents work together?

Through tracked requests. When one agent needs something from the other, the request is tracked until it is answered. If no reply comes, the request is sent again, and after that it is raised on the project where a person will see it. No request is dropped without someone seeing it.

Each agent has its own memory, and both share the team’s wiki. The agents write notes deliberately, not automatically, because a few chosen notes are more useful than a pile of incidental ones.

Why does each task load only part of an agent’s instructions?

To keep the agent focused on the task in front of it. An agent answering a quick question does not need its full guidance for setting up a new project, so each kind of task loads only the steps that apply to it. Less in the context means faster, more focused answers.

What can’t the agents do?

Change the firm’s records, see outside conversations, interpret real results, or contact clients.

The agents can read the firm’s project boards and conversations so they stay informed. They cannot change them. That limit is enforced by the software.

Conversations that include people outside the firm are excluded entirely. The epidemiologist works only with aggregated, de-identified or synthetic data, and interpreting real results is reserved for a named scientist.

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.

What should an expert team look for in AI agents?

Three things: work that can be checked quickly, a way of working the team can read and change, and a clear line where the agent’s role ends. Both of Apollo’s agents finish in the same place, with the work ready for a person to take forward. More on Apollo.