Execution Monitoring in Planning: A Comprehensive Guide for AI Agents

Introduction

As no plan can cover all possible contingencies, the ability to detect failures during plan execution is crucial to the robustness of any autonomous system operating in a dynamic and uncertain environment [2]. Execution monitoring provides this capability, serving as the bridge between planning and real-world action.

Execution monitoring is the process of continuously observing a running system to ensure that the executing plan remains valid—verifying that actions complete successfully, preconditions are met, and the system is making acceptable progress toward its goals [6]. Without effective monitoring, agents cannot reliably adapt when things go wrong.

This guide explores the foundations of execution monitoring in planning, examining core concepts, key techniques, and practical frameworks for building robust monitoring systems.

What Is Execution Monitoring?

Execution monitoring is the component of a planning system responsible for tracking plan execution and detecting when things go wrong. It serves several critical functions [3]:

Detecting execution failures by cross-referencing the actual state of the world against expected states [6]
Monitoring for serendipitous conditions where certain actions may no longer be necessary and can be skipped [6]
Triggering replanning when failures are detected [6]
Providing real-time visibility to human operators through streaming execution events [1]

In a typical execution monitoring system, the monitor receives observations of the world state and compares them against the expected state derived from the plan. If discrepancies are detected, the system can trigger recovery actions or replanning [11].

Core Components of Execution Monitoring

1. Plan Representation

Execution monitoring requires a plan representation that captures not just the sequence of actions but also the conditions under which the plan is expected to succeed. This includes [6]:

Preconditions: Conditions that must hold before an action can execute
Effects: Changes to the world state that are expected to result from actions
Causal links: Dependencies between actions that establish and require conditions
Temporal constraints: Durations, deadlines, and ordering relationships

2. State Monitoring

State monitoring tracks the evolution of the world state during plan execution. Rather than monitoring all possible conditions, effective monitors focus on causal links—the conditions that are actually relevant to plan success [6].

As one researcher noted, "our approach is specifically interested in the explicit model of time in temporal plans and in the extraction of causal links independent of the planning algorithm used" [6]. This allows the system to "spurious or irrelevant conditions and changes in the world to be ignored" [6].

3. Temporal Monitoring

Temporal monitoring tracks whether actions are executed within their specified time windows. This includes monitoring [8]:

Duration constraints: Whether actions complete within allowed time bounds
Deadlines: Whether time-critical conditions are met
Coordination timing: Whether synchronized actions occur as scheduled

In the SUO system, for example, the EA Manager monitors "several types of time constraints and monitoring a specified location for activity" [8].

4. Error Detection and Diagnosis

When a discrepancy is detected, the system must determine whether it constitutes a failure requiring intervention. Key distinctions include [6]:

Expected vs. unexpected changes: Some deviations may be anticipated and handled
Relevant vs. irrelevant changes: Not all deviations threaten plan success
Recoverable vs. unrecoverable failures: Some failures can be corrected, others require replanning

Key Frameworks and Approaches

GenPlanX: Integration with Classical Planning

GenPlanX integrates LLMs for natural language-based description of planning tasks with a classical AI planning engine, alongside an execution and monitoring framework [3]. Its architecture includes an Execution module that "carries out the actions specified in the plan" and "monitors the execution process, checking for successful completion of each action. In the event of execution failures or when new goals appear by the execution of previous actions, GenPlanX can replan to generate a new plan" [3].

This hybrid approach demonstrates how classical planning's formal guarantees can be combined with modern AI capabilities.

Plan Executive Systems

A plan executive is responsible for executing a plan and monitoring its progress. As described in one framework, a plan executive must [6]:

Choose specific timing values for each action and dispatch them at those times
Continuously monitor the plan to ensure it is evolving properly

The executive uses a scheduler and dispatcher for timing, and an execution monitor for state constraints [6].

Logic-Based Monitoring

A temporal logic-based monitoring approach uses "formulas in an expressive temporal logic specify the desired behavior of a system and its environment" [2]. During plan execution, "an execution monitor subsystem detects violations of monitor formulas in a timely manner using a progression algorithm on incrementally generated partial logical models" [2].

This approach has been "integrated on a fully deployed autonomous unmanned aircraft system" with "extensive empirical testing...using a combination of actual flight tests and hardware-in-the-loop simulations" [2].

LLM and VLM Grounding for Detection and Recovery

DoReMi is a language model grounding framework that "enables immediate Detection and Recovery from Misalignments between plan and execution" [11]. The approach leverages "LLMs to play a dual role, aiding not only in high-level planning but also generating constraints that can indicate misalignment during execution. Then vision language models (VLMs) are utilized to detect constraint violations continuously" [11]. Experiments on "various complex tasks including robot arms and humanoid robots demonstrate that our method can lead to higher task success rates and shorter task completion times" [11].

Execution Monitoring in Multi-Agent Systems

Execution monitoring in multi-agent systems presents unique challenges. In these environments, "most actions are performed by external agents, usually humans, and the monitor has no access to the state of its executing agents" [12]. This requires "different monitoring techniques, as the executor must use incoming messages to determine the status of agents and activities" [12].

In such distributed settings, the monitor must handle [12]:

Higher rates of incoming data than autonomous robot systems
Customized monitoring of each action to generate appropriate alerts
Human-agent interaction requiring alerts tailored to human cognitive capabilities

Continual Planning: Monitoring as Part of a Closed Loop

Continual planning represents the tightest integration of planning, execution, and monitoring. In this paradigm, planning is not a one-time activity but an ongoing process that occurs throughout execution [5].

As one research group described, "it is often more practical in such environments to enable agents to actively extend their knowledge as part of their plans and then revise their decisions in light of these update" [5]. The algorithm "deliberately postpones parts of the planning process to later stages in an agent's plan-act-monitor cycle and automatically determines when to switch back to refining or revising a partly executed plan" [5].

This approach requires careful design of the monitor to answer questions like [5]:

How can a planner decide which parts of the problem-solving process to postpone?
Can agents plan their later replanning?
How are early plans that include knowledge-gathering actions related to their later revisions?
How does a planner realize which kind of knowledge gathering is necessary in the first place?

Best Practices for Execution Monitoring

Focus on relevant conditions: Monitor only conditions that are relevant to plan success, not all possible world states [6]
Extract causal links independent of planner: Generate monitor conditions from plans without relying on a specific planning algorithm [6]
Support both state and temporal monitoring: Track both what happens and when it happens [6]
Design for indirect execution: In multi-agent systems, monitor through messages rather than direct state access [12]
Provide real-time visibility: Enable operators to view streaming execution events [1]
Integrate monitoring with replanning: When failures are detected, trigger appropriate recovery actions [3]

Related Concepts

Adaptive Planning Systems — Systems that modify plans in response to changing conditions
Dynamic Task Planning — Adapting plans in response to changing conditions
Constraint-Based Planning — Planning within resource and time limitations
Long-Horizon Planning — Strategies for extended multi-step task execution
AI Agent Architecture — Foundation Agent, Core Components, Agent Systems

Conclusion

Execution monitoring is essential for any AI agent operating in dynamic, uncertain environments. Without monitoring, plans are blind—agents cannot detect when they go off course or adapt to changing conditions.

As one foundational paper noted, "the ability to detect failures during plan execution is crucial to the robustness of any autonomous system operating in a dynamic and uncertain environment" [2]. The field has evolved from early systems like PLANEX on the Shakey robot to modern frameworks that integrate LLMs with classical planning, enabling natural language interaction and robust execution monitoring [3].

Effective execution monitoring requires careful attention to what conditions to monitor, how to monitor them, and how to respond when deviations are detected. The best systems focus on relevant conditions, support both state and temporal monitoring, and integrate monitoring with replanning to enable continuous adaptation [6].

As AI agents increasingly operate in complex, unpredictable environments, execution monitoring will remain a critical capability—the bridge that connects planning to successful action.

References

CodeFrame. [Phase 3] Execution Monitor View - Real-time Agent Monitoring. GitHub. 2026.
Kvarnström, Jonas, et al. A Temporal Logic-Based Planning and Execution Monitoring System. ICAPS. 2008.
Borrajo, Daniel, et al. GenPlanX: Generation of Plans and Execution. arXiv. 2025.
Argenziano, Francesco, et al. Defining and Monitoring Complex Robot Activities via LLMs and Symbolic Reasoning. arXiv. 2025.
Brenner, Michael and Nebel, Bernhard. Continual Planning and Acting in Dynamic Multiagent Environments. ACM Digital Library. 2006.
Levine, Samuel. Causal-Link Execution Monitoring for Temporal Plans. MIT. 2007.
Berry, Pauline. Interactive Execution Monitoring of Agent Teams. Journal of Artificial Intelligence Research. 2003.
Wilkins, David, et al. SUO Execution Monitoring. CMU School of Computer Science. 2003.
Zahrádka, David, et al. A Holistic Architecture for Monitoring and Optimization of Robust Multi-Agent Path Finding Plan Execution. arXiv. 2025.
CiNii Research. New HTN Planning Framework for Agents Operating in Dynamic Environments. 2025.
DoReMi. DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment. IROS. 2024.
Berry, Pauline. Related Work: Execution Monitoring. CMU School of Computer Science. 2003.

Search This Blog

Agentic Mind