Technical Summary: Agent Primitives

Problem Statement

Existing Large Language Model (LLM)-based Multi-Agent Systems (MAS) face two primary limitations:

1. Lack of Reusability and High Complexity: Current MAS architectures are often highly task-specific, relying on manually crafted agent roles and interaction prompts. This leads to increased architectural complexity and limited reusability across different tasks. Systems are tightly coupled to specific interaction patterns, causing complexity to scale with task complexity rather than being amortized through modular reuse.
2. Communication Bottlenecks: Most MAS rely on natural language for inter-agent communication. In long-context, multi-stage interactions, this medium is vulnerable to error accumulation and instability. As conversation histories grow, natural language communication accuracy degrades, and systems become highly sensitive to accumulated noise, creating a fundamental bottleneck for robust collaboration.

Methodology

The authors propose Agent Primitives, a framework that abstracts multi-agent computation into a set of reusable, latent building blocks. Inspired by neural network design (e.g., residual blocks and attention heads), the approach decomposes complex MAS architectures into recurring internal computation patterns.

Core Components

1. Agent Primitives: Three representative primitives are instantiated, each encapsulating a specific internal computation structure while exposing a standard LLM agent interface:

   • Review: An iterative self-critique operator involving a Solver and a Critic connected via a latent feedback loop.
   • Voting and Selection: A consensus operator where multiple Solvers generate candidates, and a Selector aggregates their latent representations to compute a final output.
   • Planning and Execution: A decomposition operator where a Planner generates a structured latent plan, and an Executor performs task-specific reasoning conditioned on this plan.

2. Latent Communication via KV Cache: Unlike traditional MAS that exchange natural language, Agent Primitives communicate internally via Key-Value (KV) caches.

   • Mechanism: The KV cache of one agent (representing its processed sequence) is concatenated with the system prompt of the next agent.
   • Positional Re-encoding: To maintain compatibility with Rotary Positional Encoding (RoPE), the KV states from the preceding agent are re-indexed by an offset corresponding to the length of the subsequent agent's system prompt.
   • Benefit: This avoids repeated text decoding, mitigating information degradation and noise accumulation in long-context settings.

3. Automated System Construction:

   • Organizer: An LLM-based component that automatically selects and composes primitives based on the input query. It treats primitives as reusable building blocks rather than designing task-specific agent roles.
   • Knowledge Pool: A lightweight database storing previously observed queries paired with effective MAS configurations (derived from frameworks like Multi-Agent Debate, Self-Refine, etc.). The Organizer uses this pool to guide structural decisions, abstracting retrieved systems into primitive compositions.

Key Contributions

• Abstraction of MAS: The paper introduces a task-agnostic abstraction for multi-agent computation, moving away from manual role assignment to a modular system of reusable primitives.
• Latent Communication: It demonstrates that KV-cache-based communication is superior to natural language for internal agent coordination, significantly reducing information degradation and noise sensitivity.
• Automatic Composition: The framework enables the automatic construction of MAS without manual system design, guided by a knowledge pool of successful configurations.
• Plug-and-Play Interface: Each primitive exposes the same external interface as a standard LLM agent, allowing for flexible composition and substitution.

Experimental Results

Experiments were conducted on eight benchmarks (Math, Code Generation, Q&A) using five open-source LLM backbones (Qwen3 family, DeepSeek distillation models).

• Accuracy: Primitives-based MAS improved average accuracy by 12.0–16.5% over single-agent baselines. They consistently outperformed text-based MAS (TextMAS) and showed more stable performance across different model backbones compared to other latent communication methods (e.g., LatentMAS, which degraded significantly on LLaMA-based models).
• Efficiency: Compared to text-based MAS, the proposed method reduced token usage and inference latency by approximately 3×–4×.
• Overhead: Relative to single-agent inference, the method incurred only a 1.3×–1.6× overhead in terms of token usage and latency, offering a favorable trade-off between performance gains and computational cost.
• Ablation Studies:
  • Removing the Knowledge Pool led to performance drops of 5–12%, confirming its value in guiding composition.
  • Disabling RoPE re-encoding caused severe performance degradation, especially on LLaMA-based backbones, highlighting the necessity of positional re-encoding for KV cache concatenation.
  • KV-cache communication consistently outperformed text-based communication within individual primitives.

Significance and Claims

The paper claims that Agent Primitives provide a scalable and task-agnostic foundation for building multi-agent systems. By decomposing MAS into reusable latent building blocks and utilizing KV-cache communication, the approach:

• Reduces architectural complexity and design overhead.
• Alleviates communication degradation in long-context settings.
• Enables systematic and principled development of MAS that are robust across different tasks and model backbones.

The authors position this work as a step toward reducing engineering overhead and improving the robustness of multi-agent systems, drawing a parallel to how modular components (like residual blocks) revolutionized deep learning architecture design.