If a team of AI agents works well under one model, will it still work when you swap in a different model?

This explores whether an organizational design for a multi-agent system keeps working when you swap the underlying model, or whether the design is quietly tied to the model it was tuned on. The corpus's short answer is: less than you'd hope. The clearest evidence comes from ablation studies on RACI-style accountability tags, which label which agent is Responsible, Accountable, Consulted or Informed for each task. Those tags change outcomes only when the collaboration protocol actually routes the deliverable through the accountable agent. Even then, the best placement varies by model family, so the authors conclude that every model needs its own validation rather than one design that works for all Does accountability placement actually shape multi-agent outcomes?. A design that looks right on paper is really a hypothesis you have to re-test for each model.

The pattern also shows up one layer down, in the harness: the scaffolding of tools, prompts and structure wrapped around a model. Models at every capability tier are about equally good at writing useful harness edits. But how much a model gains from those edits follows an inverted U. Weak models don't invoke the harness properly, and strong models often don't follow its instructions faithfully, so mid-tier models gain the most Do stronger models always evolve harnesses better?. This is the counterintuitive part: moving an organizational design to a *stronger* model can make it work worse, because the stronger model is more likely to override the structure.

The classical multi-agent literature from before LLMs was built on the opposite assumption. MOISE+ separates an organization into structure (roles and groups), function (global plans), and permissions and obligations, so the organization can be specified independently of whoever fills the roles Can separating structure and function improve multiagent organization design?. Gaia goes further. It treats organizational rules and the environment as first-class design objects, not side effects of individual roles Should organizational rules be designed separately from agent roles? Should multiagent systems model their environment explicitly?. Those frameworks assumed agents would reliably honor the roles they were given. LLMs break that assumption: each model family has its own interpretive habits. Walden Yan's critique explains part of the reason. Agents carry unwritten assumptions into everything they do, and those assumptions are exactly what an org chart can't capture Why do parallel subagents produce conflicting design decisions?.

The corpus does point to designs that transfer better. They make structure explicit and machine-readable instead of leaving it to the model to interpret. Reorganizing a codebase around an explicit map from runtime behavior to code let weaker planners match stronger models at finding the right code Can explicit behavior maps help weaker planners compete with stronger models?. That suggests structure carried in external artifacts is more portable than structure carried in prompts. Similarly, Prime Agent's layered state hierarchy is explicitly designed to separate harness failures from model failures Can external state caches let models solve harder problems?. That separation is what you need before you can even tell whether a design transferred.

The takeaway: treat an organizational design as two parts, a portable skeleton (explicit routing, external state, written rules) and a model-specific tuning layer (where accountability sits, how much structure the model will tolerate). The corpus has no head-to-head study that moves one design across many model families, so the per-model-validation result is still the strongest direct evidence.