A multi-agent blueprint can assign every AI agent clear roles and duties on paper, but do the agents actually stick to the plan?

This explores where agents stop following an organizational blueprint like MOISE+, which assigns roles, shared plans and obligations, and why that gap gets wider when the agents are LLMs. A caveat first: the collection describes MOISE+ but has no study measuring how closely real agents follow a MOISE+ specification. What it does have is a set of failure patterns from other work, and each one lines up with one of the three dimensions MOISE+ formalizes. Read side by side, they suggest the gap comes less from bad specifications than from what the specification assumes about the agents carrying it out.

MOISE+ splits an organization into three parts. Structure covers roles and groups, function covers the global plans, and the deontic part covers who is permitted or obligated to do what. Rules then tie the three together, so no agent is left without guidance and no validated plan gets lost Can separating structure and function improve multiagent organization design?. That design comes from classical agent research, where agents had explicit beliefs and goals that could be checked against the rules. The AAMAS tradition makes the contrast clear: its formal tools assume agents whose mental states you can inspect, while LLM agents act from implicit patterns Can classical agent architectures fix agentic AI accountability?. An obligation written into a MOISE+ specification has nothing internal to bind to in an LLM. The agent can only read it as text and hope to honor it.

The most revealing gap is the one the specification can't see. Walden Yan argues that parallel subagents fail because every action carries unwritten assumptions, so agents that each correctly fill their role still make conflicting choices Why do parallel subagents produce conflicting design decisions?. A role definition says what you are responsible for. It doesn't capture the many small decisions you make while doing it. The AgentsNet benchmark shows the same thing from the communication side. Agents agree on a strategy too late, adopt one without telling their neighbors, and accept whatever peers report without checking it Why do multi-agent systems fail to coordinate at scale?. The protocol exists on paper, but in practice the agents don't follow it reliably. The Gaia methodology names one more blind spot. If designers model roles and plans but not the environment the agents operate in, the system's behavior drifts away from the design Should multiagent systems model their environment explicitly?.

The more practical lesson is how newer systems close this gap: they stop expecting agents to obey the specification and enforce it from outside. OpenAI's ScopeOfAutonomy records turn permissions into concrete lists of allowed tools, cost limits and ending conditions How can autonomous agents act safely within defined boundaries?. Progent goes a step further. A solver, not the LLM, decides whether each change to a policy narrows the agent's permissions or widens them, so the agent can't quietly give itself more privileges Can an SMT solver prevent LLM policy drift during agent execution?. Stopping works the same way. Instructions in a prompt can't guarantee that an agent caught in a loop will halt, so the halt has to come from a supervisor outside the agent Can prompt alignment alone guarantee agent termination in loops?. More generally, reliability comes from moving memory, skills and protocols out of the model and into the surrounding harness Where does agent reliability actually come from?.

What you may not have expected to learn is that MOISE+'s permissions and obligations, which were written as norms agents would choose to follow, are being rebuilt today as hard constraints the system enforces. MOISE+ still gives a clear vocabulary for what an organization should guarantee. Modern agent stacks, though, are finding that the guarantees only hold when they live in code and supervisors rather than in the agents' understanding of their roles.