Agent Capsules: Adaptive Runtime Optimizes Multi-Agent LLM Pipeline Efficiency

The Challenge of Multi-Agent Pipelines#

Running N agents in parallel typically requires N separate LLM calls, consuming significant token budgets. A tempting optimization is to merge agents into fewer, compound calls—but doing so naively sacrifices quality through tool loss and prompt compression artifacts.

Agent Capsules Framework#

Researchers at arXiv introduce Agent Capsules, an adaptive execution runtime that treats multi-agent pipeline orchestration as a constrained optimization problem. Rather than rewriting merged prompts to compensate for compression, the framework implements three execution strategies arranged in an escalation ladder:

1. Standard mode: merged compound execution (most token-efficient)
2. Two-phase mode: intermediate routing strategy
3. Sequential mode: per-agent dispatch (highest quality, highest cost)

The runtime continuously monitors output quality using rolling-mean metrics and gates every mode switch based on empirical thresholds—automatically routing to fine-grained execution whenever quality would degrade.

Key Findings#

A controlled experiment shows that injecting more context into merged calls actually worsens compression rather than relieving it. The framework recovers quality by escalating toward finer granularity instead.

Performance on real pipelines:

• 14-agent competitive intelligence pipeline (LangGraph baseline): 51% fewer tokens in fine mode, 42% fewer in compound mode, with +0.020 and +0.017 quality gains
• 5-agent due diligence pipeline (DSPy baseline): 19% fewer tokens than uncompiled DSPy at parity, 68% fewer than MIPROv2 at +0.052 quality improvement

Notably, the framework delivers efficiency gains even before compound mode activates, through automatic policy resolution, cache-aligned prompts, and topology-aware context injection—all without per-model configuration or training data.