One AI system reviews a code change, a second writes the tests for it. Both do good work and still miss each other. The tests may refer to an older revision or check something else. Running at the same time does not resolve the dependency. A product for such workflows needs clear assignments, clear ownership and a proper exchange of results.
Collaboration needs a reason
An agent here is an AI program that works through a task with tools over several steps. Whether a second agent helps depends on the task. Independent research runs well in parallel. Two programs changing the same short passage at once mostly produce coordination. We split work when the result gets better or it genuinely saves time.
First we look at the dependencies. What can start right away? What waits on an earlier decision? What needs a view of the whole? These questions are familiar from working with people. With AI systems the answers have to be more explicit. Shared understanding does not appear by itself.
LeanCTX contains documented building blocks for such workflows. Participating agents can register, exchange messages, take assignments and share context. That makes participants and handovers addressable. It is not a guarantee of good collaboration. Splitting the work sensibly and checking the combined result stay separate jobs. [1]
Our measure is not the number of programs running at once. We care about the time to a result that actually satisfies the task. That includes writing the assignments, reading the answers and putting them together. Doing a small task directly can be the fastest route. Collaboration is a means, not a goal.
A good handover gives the next participant enough to continue responsibly.
Handovers with clear contents
A usable assignment names the goal, the permitted scope and the expected shape of the result. It also says how completion will be recognised. “Check the change” leaves too much open. “At this revision, check whether the three named failure cases are caught, and report the test results” is a clear frame. The contribution can then be checked against it.
When one participant finishes, “done” is rarely enough. The next one needs the result, the important sources, the open points and the revision everything refers to. A full history is not required. A short, ordered handover works if it keeps the essential decisions and limits. The amount follows from the next task.
Context packages are one documented way in LeanCTX to bundle selected material. The idea behind them does not depend on the format. The recipient should get something they can work with. That includes knowing whether the information is current and what they may pass on. A technically successful transfer does not answer that. [2]
Shared files need clear ownership. If two programs change the same area at once, they overwrite each other or start from different states. So we prefer bounded assignments and an explicitly assigned integration step. When ownership changes, the current state travels with it. That keeps it visible who is working on what.
Putting it together is work of its own
Even careful individual contributions can contradict each other. A text promises a feature that does not exist. A changed interface no longer matches the program calling it. So we check the combined result against the original task again. Good parts are the precondition. They are not proof that the connection holds.
A workflow also needs a way to deal with trouble. A program finds no suitable source. A tool fails. A result stays incomplete. Those states have to remain visible. Whoever takes over needs to know what is genuinely finished. An honest report helps more than an optimistic completion message.
The same holds when ordinary programs work together. One needs a confirmed number, not an explanatory paragraph. Another needs approval before it changes anything. So our view of efficiency covers Human × AI, AI × AI and System × System. The transitions differ. Everywhere the goal, the ownership and the result have to line up.
We aim this work at the finished result. That includes counting the cost of coordination. And choosing the simple direct route when it is enough. Several AI systems pay off when their contributions complement each other and the handovers hold. A product has to make both checkable: the single contribution and the result afterwards.
From the Thinkery workshop
Our view of the work behind our products. Examples help explain the ideas. Links show which features already exist. Where we describe a goal, work is still ahead of us.
Edition of 9 September 2026