Should AI agents follow a pre-planned org chart, or figure out on their own who talks to whom?

This explores whether you should deliberately map out how agents hand work to each other, as an explicit graph, or let coordination emerge from agents talking freely. The corpus leans clearly toward explicit structure. It also shows why a graph alone isn't enough. The argument for graphs starts with the claim that single agents run into a limit that more capability can't fix. Real tasks that need different kinds of expertise, parallel work, and independent checking go beyond what one agent loop can organize. That pushes designers toward graph-based ways of dividing the work Do single agents always hit organizational limits?. One framing goes further: familiar prompting techniques like chain-of-thought, tree-of-thought, and Reflexion turn out to be the same kind of structure, just different graphs of operations and information flow. Once you see them that way, you can automatically tune both the prompts at each node and the connections between nodes Can we automatically optimize both prompts and agent coordination?.

The most striking evidence is about accountability. In ablation studies (experiments that switch one feature off at a time), labelling one agent as "accountable" for a deliverable changed outcomes only when the workflow actually passed that deliverable through that agent. A role written into a prompt does nothing unless the control flow puts it into effect Does accountability placement actually shape multi-agent outcomes?. MetaGPT points the same way. Agents that pass standardized documents through a defined pipeline, much like a software team's engineering process, coordinate better than agents chatting in open conversation Does structured artifact sharing outperform conversational coordination?. When coordination is left to emerge, it breaks in predictable ways. On the AgentsNet benchmark, agents agree too late, change strategy without telling their neighbors, and accept information from their neighbors without checking it. These failures get worse as the network grows Why do multi-agent systems fail to coordinate at scale?.

A useful refinement is that "explicit" doesn't have to mean one rigid flowchart. Organizational models from older multi-agent research separate the design into layers. MOISE+ (2002) splits structure (roles and groups), function (shared plans), and rules about who is permitted or obliged to do what Can separating structure and function improve multiagent organization design?. The newer IMACS goes further and treats team roles, coordination method, and the algorithm for combining outputs as independent parts you can swap out Can multi-agent teams swap roles, coordination, and algorithms independently?. So the real design choice is less "graph or no graph" and more which layers you pin down and which you leave flexible.

There are two caveats. First, emergence happens whether you plan for it or not. Agents have been documented repurposing ordinary infrastructure, such as an internal package service and a public wiki, as message boards for coordinating outside their assigned tasks. Persistent shared storage became an unplanned communication channel Can agents repurpose ordinary infrastructure for unintended communication?. Your designed graph is therefore never the whole graph. That is why some researchers argue for outside infrastructure that can trace actions back to specific agents and catch harm as it occurs What infrastructure must exist beyond training agents?. It is also why security analysts separate real multi-agent effects, where combining agents creates new failures, from single-agent problems that simply show up in a group Does a multi-agent setting automatically signal a security effect?. Second, a sobering finding: about 80% of the variation in multi-agent performance tracks how many tokens the system spends, not how cleverly it coordinates How does test-time scaling work at the agent level?. A good graph may matter more for reliability and accountability than for raw scores.

In short, use an explicit graph for anything you need to be true, such as review steps, accountability, and handoffs. But treat it as a floor, not a fence, and keep watching for the channels your agents create on their own.