CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

CEPO is a framework that splits multi-agent prompts into a shared coordinator prompt and agent-specific executor prompts, and gives the underlying prompt optimizer role-specific meta-instructions for each layer, improving four existing optimizers across 11 benchmarks in four domains.

Multi-agent systems need both system-level coordination and agent-level execution, but existing prompt optimizers for MAS update only agent-specific prompts in which these two functions remain coupled. Because coordination rules live inside each agent prompt, revising them requires consistent changes across multiple prompts with different execution responsibilities. The paper asks how prompt structure and optimization guidance can make existing prompt optimizers more effective for MAS.

CEPO decomposes the prompt collection into a shared coordinator prompt p_C and agent-specific executor prompts p_E^(i), where the effective prompt for agent A_i is p_C prepended to p_E^(i). Role-aligned optimization supplies two meta-instructions to the underlying optimizer: Φ_C directs coordinator updates toward team-level collaboration such as cross-agent communication and conflict handling, while Φ_E directs executor updates toward task scoping, evidence use and concrete procedures. Both layers are still selected by the same system-level validation objective J; the meta-instructions provide different guidance but not separate numerical objectives. The default procedure is sequential: Phase 1 optimizes the coordinator for T_C iterations with initial executors fixed, and Phase 2 optimizes the executors of a user-specified subset I for T_E iterations with the optimized coordinator fixed, selecting prompts by validation score. CEPO specifies prompt structure and guidance while the underlying optimizer performs candidate generation and search.

• Applying PE2 + CEPO to MDocAgent improves every one of the five document benchmarks, with an average gain of 4.3 points and MMLongBench-Doc rising from 54.1 to 61.0. §4.2
• Applying PE2 + CEPO to DocAgent gives an average gain of 2.7 points, with the largest gain on MMLongBench-Doc rising from 59.5 to 67.4. §4.2
• CEPO on top of MASPO beats both the baseline and MASPO alone on all five mathematical reasoning and code generation benchmarks, with an average of 73.89 versus a baseline of 68.82. §4.2
• Component analysis on MMLongBench-Doc: an LLM-generated coordinator without optimization raises 54.1 to 55.6, vanilla PE2 executor-only optimization reaches 57.9, coordinator-only optimization reaches 58.6, and CEPO reaches 61.0. §4.2
• CEPO with T_C = T_E = 1 reaches 60.1, exceeding vanilla PE2 with three executor iterations by 2.2 points while using fewer optimizer and MAS API calls, whereas vanilla PE2 with six executor iterations reaches 57.4. §4.3
• In the guidance swap experiment, coordinating guidance on the coordinator and execution guidance on the executor is best at 61.0, reversing the assignment is worst at 57.7, and using the same guidance for both layers gives intermediate results. §4.3

• The paper states that the sequential coordinator-first schedule is an implementation choice that performed best among those evaluated only in its document-system ablation, so other schedules may suit other systems.
• Document benchmarks lack dedicated training sets, so the authors partition test sets into document-level 10/10/80 train/validation/test splits using a fixed seed.
• Answers are scored with GPT-5-mini under an LLM-as-judge protocol, and the authors state they plan to release the codebase only upon acceptance.
• The ethics statement warns that easier prompt optimization may encourage deployment without sufficient validation and that optimized prompts can still produce unintended behavior in high-stakes or out-of-distribution settings.

• The coordinator prompt is shared across all agents and prepended to each agent's executor prompt, so a single coordination update applies consistently while execution instructions stay separately adjustable. §3.2
• A coordinator update changes the shared instructions received by all agents, whereas an executor update changes only the local instructions of agent A_i. §3.2
• Role-aligned optimization uses meta-instructions Φ_C and Φ_E; both prompt layers are selected according to the same system-level objective J, so the guidance differs but no separate numerical objectives are defined. §3.2
• The default schedule optimizes the coordinator first with initial executors fixed, then optimizes executors of a subset I; I = ∅ gives coordinator-only optimization and I = {1,...,N} permits optimizing every executor. §3.3
• All schedules in the schedule ablation use the same total budget of six iterations; sequential coordinator-first scores 61.0 and executor-first 60.7, while alternating gives 60.1 and 58.2. §B.2
• With one to five agents, unoptimized MDocAgent gains little from extra agents and vanilla PE2 falls below the unoptimized result at three agents, while PE2 + CEPO increases monotonically to 61.0 and its advantage grows from 3.9 points at two agents to 6.9 at five. §4.3
• CEPO improves every optimizer-benchmark combination with gains of 0.6 to 6.5 points over PE2, ProTeGi and OPRO, and also improves MASPO, a multi-agent prompt optimizer. §4.3
• Extending vanilla PE2's prompts to a comparable length does not recover CEPO's performance: PE2 (long) with 1410 tokens scores 54.0 while PE2 + CEPO with 1408 tokens scores 61.0. §4.3