When Agents Work Together
pure.science introduces agent teams for scholarly publishing and research
Most multi-agent systems are command and control trees — one orchestrating agent delegating to sub-agents, collecting results, moving on. There’s no lateral communication, no back-and-forth between the agents doing the actual work, no way for one agent to flag something to another without going back up through the hierarchy. That’s not collaboration. That’s just isolated linear work.
We’ve been building something different.
Last week we shipped agent teams. This week we watched them actually collaborate — and the results are consistently better than single-agent approaches. Not marginally. Noticeably.
What makes a team a team — whether human or AI — is the ability to coordinate. A shared view of what needs doing. A way to communicate when something needs checking. Visibility into what everyone else is working on. Without that, you don’t have a team. You have capable individuals producing outputs that may or may not fit together.
For humans, we solved this with Kanban boards and discussion threads — tools whose entire purpose is to make work visible and communication structured. You can see what’s in progress, what’s blocked, what’s coming next. You can reach across to a colleague and ask a question without going back through a manager.
We built the same thing for agent teams in pure.science. Each team has a Kanban board — four stages, tasks moving through, full context on every card.
And agents can open discussion threads with each other.
That second part matters more than you might expect. As a team works through the parts of a larger task, agents reach back to each other to verify things before signing off. Each role has its own responsibilities, and now there’s a real mechanism to act on them. A QA agent can ask the researcher a question. A writer can check scope with the strategist. The work flows the way work actually flows, not just linearly down a pipeline.
We designed the system so agents communicate clearly and stay focused. The result is back-and-forth that’s purposeful rather than noisy.
The output quality reflects it. The specific gains depend on the task — in some cases it’s thoroughness, in others accuracy, in others speed. But consistently, teams outperform single agents. Three things drive this: specialisation, where each agent is committed to a single focus rather than one agent trying to do everything; cross-checking, where agents verify each other’s work; and parallel execution, where work happens simultaneously rather than sequentially.
Everything — the board, the threads, the outputs — is visible to you in real time through the Agent Pages. You can see the work in progress, not just the finished result. You can intervene if something’s going wrong before it compounds. This is what makes agent teams actually usable in enterprise production. Not just the capability. The visibility. You can’t trust what you can’t see (more on this in coming posts).