At a Glance
Better memory and constraint-guided coordination cut wasted actions, making multi-agent systems run ~30% faster, use far fewer language model calls, and succeed more often.
ON THIS PAGE
Key Findings
Failed actions (trying things that don’t work) are the main reason multi-agent plans take so long — caused by poor state tracking when agents only see part of the world, and by agents stepping on each other. Mosaic fixes both with lightweight, agent-centered memory semantic-capability-matching-pattern that records object positions relative to each agent, and a constraint solver that assigns actions so agents don’t conflict Dynamic Task Routing Pattern. The result: fewer wasted steps and far fewer calls to large language models, so plans run faster and more reliably in embodied tasks.
Avoid common pitfallsLearn what failures to watch for
Data Highlights
127–32% faster end-to-end execution across AI2-THOR and search-and-rescue benchmarks
230–33% fewer large-language-model calls compared to baselines
34–10 percentage points higher task success rates across evaluated scenarios
What This Means
Engineers building multi-robot fleets or multi-agent virtual assistants will see lower run time and API costs and more reliable task completion. Technical leaders evaluating agent orchestration should consider lightweight, agent-centric memory and constraint-based action allocation to scale systems with fewer retries and less compute waste. The Dynamic Task Routing Pattern can help teams reason about how to allocate actions across agents more efficiently.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results come from simulated embodied benchmarks (AI2-THOR and search-and-rescue); sensors, noise, and unmodeled physics in the real world could reduce gains. The constraint solver adds CPU work per planning step, which is small compared to language-model calls here but may grow with many agents or extremely large action spaces. The memory design assumes geometry and relative transforms matter; problems that are purely symbolic or lack stable spatial relationships may not benefit as much. See State Inconsistency for related pitfalls and Resource Exhaustion as a potential scaling concern.
The Details
Long waits in multi-agent planning mostly come from repeated failed actions caused by two problems: agents don’t keep a consistent view of where objects are when they only see part of the scene, and independent planning makes agents produce redundant or conflicting moves. Mosaic addresses both without relying on heavier or more frequent language-model queries. Each agent maintains a compact semantic memory that stores objects in coordinates relative to the agent; those relative entries can be transformed between agents to build a consistent, lightweight world state. At each planning step, Mosaic uses an integer linear program (a simple constraint optimizer) to allocate actions across agents while enforcing physical feasibility and preventing conflicts.
On standard embodied benchmarks Mosaic reduces end-to-end execution time by about 27–32%, cuts language-model calls by 30–33%, trims plan length, and raises success rates by 4–10 percentage points. The savings come from avoiding retries and unnecessary coordination attempts: accurate local memory reduces mistaken actions, and the solver prevents agents from duplicating work or colliding. For teams building or evaluating multi-agent systems, the takeaway is that smart, small memory structures plus constraint-guided coordination give big practical wins in latency, cost, and reliability. Future work should validate these gains on real robots and explore solver scalability for very large agent pools.
Need expert guidance?We can help implement this
Credibility Assessment:
Authors are affiliated with Adobe Research (recognized industry lab) and have modest h‑indices (~7–13). ArXiv venue and low citations limit credibility to a solid mid‑level rating.