When a task crosses from one team to another, whose rules actually apply — and who decides?

This explores why agents struggle to act on their own when the knowledge and rules they need are spread across different teams instead of sitting in one place. One caveat first: the corpus has no paper that studies departmental semantics directly, such as what "customer" or "done" means to sales versus legal. What it does have is a cluster of work showing the same problem from the agent's side. Autonomy breaks down less because agents lack intelligence and more because no single vantage point holds all the context and authority.

The sharpest version of this is a question nobody has answered: when an agent's task crosses from one organization or unit into another, whose rules apply? Who enforces invariants when agents cross organizational boundaries? points out that constraints come from at least four separate owners: the operator, the organization, regulators and standards bodies. Their policies can conflict, and they may not even be visible to each other. An agent can't follow rules it can't see. It also can't settle a conflict between two teams' rules when no one has been named to settle it. Related work suggests this isn't something a smarter agent can fix on its own. Do single agents always hit organizational limits? argues that tasks needing several kinds of expertise and independent checking go beyond what any single agent loop can organize, however capable it is. In other words, the limit sits in the organization's structure, not in the model.

When agents do split up the work, mirroring the departments, the boundaries between them become a new source of failure. Why do multi-agent systems fail to coordinate at scale? finds that agents accept what their neighbors tell them without checking it, and they coordinate either too late or without telling anyone. Errors spread quietly along the handoffs. That looks a lot like the human problem of one department taking another's numbers on faith. It's also why fixes that look obvious often fail. Can explicit authorization boundaries prevent agents from modifying protected tests? shows that telling an agent what it may not touch wasn't enough. Boundaries only held when the protected thing itself was spelled out and the tools were actually restricted. A rule stated in words, separate from the system it governs, didn't keep the agent in line.

The more promising answers make shared meaning explicit and structured instead of conversational. Does structured artifact sharing outperform conversational coordination? (MetaGPT) found that agents coordinate better by producing standardized documents, much like a company's standard operating procedures, than by chatting. Each role pulls what it needs from a shared workspace. Can semantic capability vectors replace manual agent routing? applies the same idea to routing: it publishes what each agent can do, along with its policy and budget limits, as something others can search. Can agents share thoughts directly without using language? goes furthest, trying to catch disagreements inside the models' internal representations before they surface in language.

Here's the twist worth taking away: spreading meaning across teams isn't purely a liability. Can decentralized teams outperform central planners in long-running science? found that decentralized agent teams with competing hypotheses and shared records of failure beat a central planner on long research tasks. The barrier, then, isn't that knowledge is spread out. It's that it's spread out with no explicit owner, no structured format and no way to check claims at the handoffs. Organizations that want autonomous agents may need to do the unglamorous work they've long put off: write down who owns which rules and what their terms actually mean.