Improving Team Rules Separately From Agent Procedures

When an agent team produces a poor result, the defect may sit in one specialist's procedure or in how the team combines its work. Editing every prompt at once makes that distinction harder to inspect.

Imagine a hypothetical procurement team checking whether supplier quotations satisfy a purchase request. One agent reads specifications, another extracts prices and a third prepares the comparison. Under the assumed policy, the team may recommend a supplier but cannot place an order.

If the agents silently resolve conflicting delivery dates, the shared handoff rule needs attention. If only the price reader confuses a unit price with an order total, its extraction procedure needs attention. Those repairs belong at different levels even though both affect the final recommendation.

Give shared rules a separate change surface

CEPO, by Jongha Kim and colleagues, separates a shared coordinator prompt from agent-specific executor prompts. Here “coordinator” means instructions prepended to each agent, not an additional manager agent.

The optimizer receives different guidance for each layer. Shared updates address collaboration and evidence handoffs; specialist updates address how a particular role performs its task. Both are selected using the whole system's validation score, rather than separate numerical rewards for coordination and execution.

The paper evaluates this structure with four existing prompt optimizers across eleven benchmarks covering documents, mathematics, code and video. In its document-system analysis, optimizing the shared rules and one selected executor performs better than expanding updates to every executor. More edits are not automatically a stronger candidate.

Those results concern the tested systems and datasets. Document answers are scored by a language-model judge, and the paper promises a later code release. Its benchmark gains support examining the separation; they do not measure procurement performance or permission enforcement.

Review the change and the assembled team

For WTK, the useful question is how a catalog team package represents shared obligations alongside role-specific procedures when it is compiled for a target runtime. A shared-rule revision can affect every participant even when only one block of text changes.

Our component-agent review asks whether dividing work improves the finished outcome. CEPO adds a more specific improvement question: which instruction layer should change when that outcome fails?

The procurement example makes the review concrete. Did a shared update preserve disagreements for the final comparison? Did a specialist update fix price extraction without dropping delivery constraints? Did either change cause the team to imply that a recommendation was an approved purchase?

Our plugin-composition note similarly keeps component evidence separate from evidence about the assembled workflow. A locally useful instruction still needs assessment in the team that consumes it.

People can set the purchasing boundary and the criteria for an acceptable comparison, then oversee package revisions and exceptions. Separating shared rules from specialist procedures could make that oversight more precise while leaving routine collaboration to the agents.

Source: Kim and colleagues, CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems, arXiv:2610.34722v1, October 26, 2026, especially Section 3 and Appendix B.2. The procurement example is ours.