Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use

Feb 12, 2026•Hanbing Liu, Chunhao Tian, Nan An +4•View PDF

TL;DR Highlight

INTENT: a lightweight planning layer that simulates future tool call costs before execution, blocking or replanning when an LLM agent would exceed budget.

Who Should Read

Backend/ML engineers running agent systems that make multiple calls to paid external APIs (financial data, satellite imagery, legal databases, etc.). Especially useful with MCP servers where each tool call has a real dollar cost.

Core Mechanics

INTENT adds a cost simulation layer before tool execution — the agent plans a full call sequence, estimates total cost, and only proceeds if within budget
When projected cost exceeds budget, INTENT triggers replanning with explicit budget constraints rather than just blocking
Reduces overage incidents (budget exceeded) by 87% compared to reactive cost-checking (check after each call)
Planning overhead is under 200ms per agent turn — negligible compared to actual API call latency
Works with any tool-calling LLM and any cost model — just register tools with their cost estimates

Evidence

Budget overage rate: 34.2% baseline (no cost control) → 4.4% with INTENT (-87% reduction)
Task completion rate maintained at 89.3% with INTENT vs. 91.2% without — only 1.9%p trade-off for 87% overage reduction
Planning overhead: median 180ms per agent turn across 500 test scenarios
Replanning success rate when budget exceeded: 73% of replanning attempts produce a valid lower-cost plan

How to Apply

Register each tool in your agent with a cost estimate (fixed or formula-based) — even rough estimates significantly improve budget adherence
Set a per-session budget and pass it to INTENT; the layer handles the rest without modifying your agent's core logic
For tools with variable cost (e.g., cost depends on query complexity), use the upper-bound estimate to be conservative — INTENT's value is in preventing overages, not optimizing spend

Code Example

snippet

Terminology

INTENTThe paper's proposed framework — a cost-aware planning layer for LLM agents that simulates tool call costs before execution.

MCP (Model Context Protocol)A standard protocol for connecting LLM agents to external tools and data sources.

replanningWhen an agent's initial plan is rejected (e.g., too expensive), replanning generates an alternative plan that satisfies the constraints.

tool call costThe monetary or computational cost incurred each time an agent invokes an external API or tool.

Original Abstract (Expand)

We study budget-constrained tool-augmented agents, where a large language model must solve multi-step tasks by invoking external tools under a strict monetary budget. We formalize this setting as sequential decision making in context space with priced and stochastic tool executions, making direct planning intractable due to massive state-action spaces, high variance of outcomes and prohibitive exploration cost. To address these challenges, we propose INTENT, an inference-time planning framework that leverages an intention-aware hierarchical world model to anticipate future tool usage, risk-calibrated cost, and guide decisions online. Across cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts such as tool price changes and varying budgets.