TAPE: Tool-Guided Adaptive Planning and Constrained Execution in Language Model Agents

Abstract: LLM (LM) agents have demonstrated remarkable capabilities in solving tasks that require multiple interactions with the environment. However, they remain vulnerable in environments where a single error often leads to irrecoverable failure, particularly under strict feasibility constraints. We systematically analyze existing agent frameworks, identifying imperfect planning and stochastic execution as the primary causes. To address these challenges, we propose Tool-guided Adaptive Planning with constrained Execution (TAPE). TAPE enhances planning capability by aggregating multiple plans into a graph and employing an external solver to identify a feasible path. During execution, TAPE employs constrained decoding to reduce sampling noise, while adaptively re-planning whenever environmental feedback deviates from the intended state. Experiments across Sokoban, ALFWorld, MuSiQue, and GSM8K-Hard demonstrate that TAPE consistently outperforms existing frameworks, with particularly large gains on hard settings, improving success rates by 21.0 percentage points on hard settings on average, and by 20.0 percentage points for weaker base models on average. Code and data available at here.

• The paper introduces TAPE, a framework that mitigates planning and sampling errors by integrating plan graph aggregation, external solver-based path selection, and constrained decoding.
• It employs a methodical approach with mismatch-triggered replanning to ensure global consistency and robust feasibility under strict operational constraints.
• Empirical evaluations demonstrate that TAPE significantly outperforms ReAct and Plan-and-Act, achieving up to a 46.0% success rate in challenging environments.

TAPE: Tool-Guided Adaptive Planning and Constrained Execution in LLM Agents

Introduction and Problem Analysis

TAPE addresses failures in LM agents operating in environments where a single action can irreversibly preclude goal achievement due to strict feasibility constraints. These environments are prevalent in robotics, code synthesis under execution quotas, time-sensitive system control, and complex reasoning domains with tool usage or cost limits. The paper identifies two major sources of such irrecoverable failures in ReAct and Plan-and-Act (PA) agent frameworks: planning errors, resulting from imperfect or hallucinated internal rollouts and limited capacity for global consistency, and sampling errors, stemming from stochastic token generation that deviates from planned actions. As the task horizon ($T$) grows, both failures accumulate, sharply degrading success rates.

Figure 1: Sources of irrecoverable failure in prior frameworks, toy analysis of error compounding in Sokoban, and a conceptual overview of TAPE that resolves these failures through plan graph aggregation, formal solver-based path selection, and constrained execution.

TAPE Framework

TAPE (Tool-guided Adaptive Planning with constrained Execution) explicitly targets both error modes by restructuring the agent pipeline into modules: (1) plan graph construction, (2) external solver-based path selection, (3) constrained execution, and (4) mismatch-driven replanning.

Figure 2: Architectural overview of TAPE, detailing plan graph construction from multiple LLM-sampled abstract trajectories, solver-based cost-aware path selection, constrained decoding for execution, and mismatch-triggered replanning.

Plan Graph Construction

Multiple diverse plans are sampled from the LM's internal model at each step. These trajectories are abstracted and merged by folding equivalent states (using a state projection and merging function) to form a compacted plan graph. This aggregation increases the diversity of potential viable actions at each decision point, exponentially reducing per-step planning error compared to single-trajectory execution.

Path Selection via External Solver

Nodes and edges in the plan graph are annotated with predicted rewards and costs, leveraging the LM's internal models for reward and cost estimation. A discrete path selection problem, formalized as an ILP, is solved over the plan graph, subject to hard feasibility constraints (budget, tool limits, etc). This process ensures that only fully feasible plans are selected—unlike LM generation alone, which cannot enforce formal constraints globally.

Constrained Execution

During execution, TAPE enforces constrained decoding, restricting the LM to only emit the next planned action as selected by the ILP trajectory. This mechanism directly suppresses sampling error, eliminating stochastic or drift-induced action deviations that might violate constraints or lose progress.

Mismatch Check and Replanning

The agent continually monitors congruence between planned and realized states. Any detected divergence (environmental stochasticity, unanticipated tool result, or model hallucination) triggers a full replanning cycle: sample new plans, rebuild the graph, and rerun path selection. This loop substantially improves robustness to all observed agent-environment mismatches.

Theoretical Guarantees

The formal analysis distinguishes TAPE from previous frameworks using bounds on compounded error. For a minimal task horizon $T$, ReAct success decays as $$U_\mathrm{ReAct} = [(1-\epsilon_p)(1-\epsilon_s\delta_b) + \epsilon_p\epsilon_s\delta_r]^T$$, where $\epsilon_p$ and $\epsilon_s$ are planning and sampling error rates, and $\delta_b$/$\delta_r$ quantify outcome viability on deviations. PA achieves $U_\mathrm{PA}\ge U_\mathrm{ReAct}$ by reducing $\epsilon_s$ via in-context plan following, but planning errors remain. TAPE further decreases planning error exponentially as the action candidate set size per state increases; with $d$ candidate plans per node, the upper bound becomes $$U_\mathrm{TAPE} = \prod_{t=0}^{T-1} [1-(\epsilon_p)^{d(v_t)}]$$, and with constrained decoding, the effective sampling error is eliminated ($\epsilon_s \approx 0$). Thus, TAPE's guarantee strictly dominates prior agent frameworks.

Empirical Evaluation

TAPE is evaluated on four agentic environments (Sokoban, ALFWorld, MuSiQue, GSM-Hard) under irrecoverability-inducing constraints (tight budget, limited tool calls, etc), using multiple LLMs including gpt-4.1-mini up to gpt-5-nano.

Figure 3: TAPE consistently achieves higher success rates than ReAct and Plan-and-Act across benchmarks and task difficulties, with pronounced gains on harder settings and for weaker LMs.

Ablation on Sokoban with a known oracle identifies sharp reductions in both planning and complete elimination of sampling errors with TAPE. Notably, TAPE achieves a 46.0% success rate, compared to 5.0% (ReAct) and 17.0% (PA). Across tasks, gains range from $+21$ percentage points on hard settings and $+20$ points for weak base models—large, robust improvements. TAPE maintains its margin as the per-environment action budget increases, directly translating available slack into higher success, in contrast to the plateau observed with other agents.

Impact of Model Capacity and Design Parameters

Figure 4: Success rate as a function of LM backbone; TAPE’s gains are especially pronounced for weaker LMs, confirming that formal plan aggregation and constrained execution compensate for reduced LM planning ability.

Figure 5: TAPE success rate increases monotonically with larger step budgets, forking above other frameworks, showing true exploitation of recoverability.

The framework’s sensitivity to the number of aggregated plans ($M$) is moderate, with $M=4$ typically optimal; excessive values introduce noise due to LM-internal inconsistencies in graph construction, affirming a trade-off between diversity and reliability.

Component Ablation

All three pillars—external solver, constrained execution, and replanning—are empirically essential. Disabling constrained execution triggers the sharpest decline in success, exposing the non-negligible nature of stochastic LLM rollouts, while forgoing the solver or replanning also undermines feasibility under constraint drift and environment deviation.

Implications and Future Directions

Pragmatically, TAPE advances agent robustness for any constrained interaction regime where prior approaches falter due to error compounding and stochastic execution. Theoretically, it establishes a route to combine neural predictive models with symbolic, globally-consistent planning and actuation. Extensions include dynamic or learned plan graph construction, solver-agnostic pipeline design, and integration with continuous control or open-world simulation. Future work should focus on improved automated graph construction that more accurately merges state abstractions, adaptive solver selection, and more general cost-modeling for less structured environments.

Conclusion

TAPE presents a principled agent framework mitigating both planning and execution errors in LM agents under irrecoverability and feasibility constraints. By aggregating diverse rollouts, leveraging solvers for constraint satisfaction, enforcing action determinism, and adaptively replanning, TAPE delivers strong empirical performance and establishes new standard success rates across demanding agentic tasks and model regimes. This formal-neural hybrid method offers a convergent direction for robust, deployable LM agents in high-stakes and resource-constrained applications.