Inside AI agent systems, does who's labeled 'in charge' matter less than how work actually gets routed between agents?

This explores whether, in agentic AI, deep structure (how work is routed, what scaffolding surrounds the model, what infrastructure sits underneath) decides outcomes more than surface features like role labels, model size, or clever agent designs. If you meant market incumbents versus startups, the notes retrieved here don't address that directly. They do say a lot about structure beating surface inside the systems themselves, and the answer is a fairly strong yes.

The clearest example comes from experiments on accountability. Tagging an agent as the one 'accountable' for a deliverable changes nothing unless the workflow actually passes that deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. The label is surface and the routing is structure, and only the structure counts. The same pattern shows up in reliability research. Agents become dependable by moving memory, skills, and interaction protocols out of the model and into a surrounding 'harness', not by getting smarter models Where does agent reliability actually come from?. A blunter result: about 80% of the variation in multi-agent performance comes from how many tokens the system spends, not from how cleverly the agents coordinate How does test-time scaling work at the agent level?. Much of what looks like intelligent teamwork turns out to be budget.

Structure doesn't mean rigid hierarchy, though. In a 25,000-task experiment, the winning setup fixed the order of work from outside but let agents choose their own roles. It beat top-down control by 14% and fully self-organizing teams by 44% Do self-organizing agent teams outperform rigid hierarchies?. Related work finds that a single agent hits organizational limits no amount of extra capability or context can fix Do single agents always hit organizational limits?. Scale exposes those limits too. Coordination breaks down predictably as networks grow, because agents accept information from neighbors without checking it Why do multi-agent systems fail to coordinate at scale?.

Here's the twist you might not have expected. If structure matters more than raw model power, the biggest models lose their built-in advantage. Small language models can handle most agent subtasks at 10–30× lower cost, so the sensible design uses small models by default and calls large ones only when needed Can small language models handle most agent tasks?. That disrupts frontier-model incumbents from below. The value moves to whoever owns the orchestration layer. At the institutional level, the existing structures look weak. Agency law and principal-agent economics identify the right problems but can't enforce anything against fast, opaque agents Can traditional governance tools actually control AI agents?. That's why some researchers argue for new internet-style infrastructure that can trace actions back to agents, shape how agents interact, and detect and fix harm What infrastructure must exist beyond training agents?. Even familiar categories are bending. Executives mostly see agents as coworkers rather than tools Does agentic AI fit the tool-or-coworker framework?, and some agentic systems now hand tasks back to humans Can AI systems delegate tasks back to humans?.

So agentic systems reward structure, but not necessarily today's structural incumbents. The advantage goes to whoever controls routing, harnesses, and infrastructure, and right now those layers are being built, not inherited. For the market version of this question, look at the collection's agentic-commerce material, which this set of retrievals didn't reach.