What the strongest study shows: LLM-based planners can handle long-running tasks
The most direct evidence comes from a 2024 study that tested an LLM-based multi-agent planner (LLaMAR) on long-horizon tasks in partially observable environments. In household tasks and search-and-rescue simulations, LLaMAR achieved a 30% higher success rate than other state-of-the-art LM-based multi-agent planners [2]. This means that for every 10 tasks a competing system completed successfully, LLaMAR completed 13. The system uses a plan-act-correct-verify framework that allows agents to self-correct from execution feedback without needing a simulator or oracle — a key feature for real-world deployment where you can't rewind mistakes.
The study also introduced MAP-THOR, a test suite with tasks of varying complexity, which provides a standardized way to compare multi-agent planners. The 30% improvement is a large effect, suggesting that the cognitive architecture — not just the underlying language model — matters a lot for long-running reliability.
Safety and trust mechanisms: what makes a multi-agent system trustworthy
Trust in multi-agent systems isn't just about task completion — it's about safety and predictability. A 2022 study introduced the Distributed Simplex Architecture (DSA), a runtime assurance technique that provides safety guarantees for multi-agent systems under distributed control [3]. Each agent runs a local safety monitor, and if the system detects a risk (like an impending collision), it switches from the advanced controller to a provably safe baseline controller. The study proved safety for flocking with collision avoidance, safe navigation of ground rovers, and microgrid operation — all real-world tasks that run for extended periods.
Another 2023 study tackled trust from a formal logic perspective, developing a model-checking framework that can verify properties about trust and commitments simultaneously [4]. The key finding is that this verification remains computationally efficient (polynomial time) even as the system scales, meaning you can formally prove trust properties without exponential slowdown. This is important for long-running tasks because you need to guarantee that agents will honor their commitments over time, not just in the short term.
The catch: persistent agents create new security vulnerabilities
A 2026 study revealed a serious risk: autonomous LLM agents that run as long-lived processes with persistent memory and scheduled tasks can be exploited by worm-like attacks [5]. The researchers demonstrated zero-click autonomous propagation across three production agent frameworks, achieving 3-hop cross-platform transmission without any platform-specific adaptation. This means an attacker could inject malicious content into one agent's memory, and that content could automatically spread to other agents on different platforms, escalate privileges, and exfiltrate data — all without human intervention.
The study also proposed defenses: a "No Persistent Worm Propagation" theorem and a system called RTW that blocks write-before-exposed-read re-entry, protects static configuration files, and prevents untrusted summaries from entering trusted memory. These defenses eliminate the persistence-re-entry-action chain while preserving normal workflows. The key takeaway is that trust in long-running multi-agent systems requires not just task competence but also robust security architecture — a lesson from the cybersecurity world that now applies to AI agents.
Practical constraints: planning complexity and human expertise
A 2024 study on multi-UAV (drone) systems highlighted a practical challenge: specifying operations for multi-agent systems in dynamic real-world contexts requires specialized expertise [1]. The researchers proposed declarative languages to reduce this burden, allowing users to plan high-level multi-UAV operations without deep technical knowledge. The system was demonstrated with three simulated UAVs running as software-in-the-loop, showing automatic on-the-fly re-planning if any UAV aborts the mission. This is a step toward making multi-agent systems accessible, but the simulation-only validation means real-world performance remains unproven.
The trade-off is clear: systems that are easier to use (like the declarative language approach) may sacrifice some of the formal safety guarantees that make systems trustworthy. Conversely, systems with strong safety proofs (like the Distributed Simplex Architecture) require more expertise to set up and maintain. For long-running real-world tasks, the choice depends on the risk tolerance of the application — a drone swarm delivering packages in a city needs stronger safety guarantees than a team of cleaning robots in a warehouse.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2022 to 2026, 3 from 2024 or later, 2 in Q1 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 59 papers retrieved from a database of over 500 million.
Sources used in this answer
Towards Autonomous Multi-UAV U-Space Operation Planning
Proposes a multi-UAV planning architecture with declarative languages to reduce the expertise needed for multi-drone operations; demonstrated in simulation with three UAVs, showing automatic re-planning when missions abort [1].
Long-Horizon Planning for Multi-Agent Robots in Partially Observable Environments
LLaMAR, an LLM-based multi-agent planner, achieved a 30% higher success rate than other state-of-the-art LM-based planners in long-horizon household and search-and-rescue tasks in partially observable environments [2].
A distributed simplex architecture for multi-agent systems
The Distributed Simplex Architecture provides runtime safety guarantees for multi-agent systems under distributed control, proven on flocking with collision avoidance, rover navigation, and microgrid operation [3].
Model checking combined trust and commitments in Multi-Agent Systems
Develops a formal logic (TCTLC) and polynomial-time model-checking algorithm to simultaneously verify trust and commitment properties in multi-agent systems, implemented in the MCMAS model checker [4].
Autonomous LLM Agent Worms: Cross-Platform Propagation, Automated Discovery and Temporal Re-Entry Defense
Demonstrates zero-click autonomous worm propagation across three production LLM agent frameworks, achieving 3-hop cross-platform transmission without platform-specific adaptation; proposes formal defenses under a No Persistent Worm Propagation theorem [5].