AgentDiet: Agentic Nutrition Management

• AgentDiet is a modular, multi-agent framework for personalized nutrition management that leverages specialized agents and knowledge graphs to ensure dietary adequacy and budget compliance.
• It employs advanced optimization formulations, including linear and mixed-integer programming, to dynamically adapt meal plans under price shocks and nutritional constraints.
• The framework features inference-time trajectory reduction for LLM-based agents, significantly cutting input tokens and computational costs while maintaining high accuracy.

AgentDiet encompasses a class of agentic frameworks, architectures, and algorithms for personalized nutrition management and dietary recommendation, distinguished by their use of modular multi-agent systems, inference-time efficiency mechanisms, optimization-driven planning, and integration of knowledge graphs and user data. Architectures branded "AgentDiet" address domains including household budget-aware meal planning, dietary supplement advice, behavioral barrier identification, multimodal intake logging, and context-efficient LLM trajectory management. The umbrella term covers both consumer-facing conversational agents and backend frameworks for continuous, personalized dietary optimization.

1. Multi-Agent System Design and Workflow

AgentDiet systems are structured as modular, multi-agent environments in which specialist agents communicate via shared databases, event-driven message buses, or learned graph topologies (Syed et al., 24 Dec 2025, Xu, 8 Jan 2026, Shi et al., 10 Oct 2025). Typical agent roles include:

• Budget Agent: Aggregates household income and fixed expenses, computing the weekly food budget $B_w$.
• Price Monitor Agent: Ingests live supermarket data and detects price shocks via $\bigl|p_{i}(t)-p_{i}(t-1)\bigr|/p_{i}(t-1)>\tau$ events.
• Nutrition Agent: Maintains a nutrient database $\{c_{i,n}\}$ and enforces the nutritional adequacy constraint $\sum_i c_{i,n}x_i \ge R_n$.
• Health Personalization Agent: Adjusts $r_{m,n}$ per member based on health metadata; sets $R_n=\sum_{m}r_{m,n}$.
• Cultural Preference Agent: Encodes dietary constraints (halal, vegetarian, fasting, allergy) and enforces exclusions $x_i=0$ for $i\in I_{\text{excluded}}$.
• Substitution Agent: Utilizes a substitution graph $G=(V,E)$ with cost-nutrient weighted edges $w_{ij}$ for dynamic replacement of food items.
• Procurement Agent: Translates optimized dietary plans $\{x_i\}$ into operational shopping lists.
• Explainer Agent: Logs key decision steps and generates human-readable justifications.

Agents operate over a central knowledge base, which stores ontologies, substitution graphs, household/user profiles, and health metadata. Orchestration is predominantly event-driven, e.g., upon detection of a price shock, the Substitution Agent is triggered, resulting in a re-solved constrained linear program for dietary planning (Syed et al., 24 Dec 2025).

2. Optimization Formulations for Dietary Planning

AgentDiet frameworks formalize dietary recommendation as a constrained optimization problem, often adopting linear or mixed-integer programming modalities (Syed et al., 24 Dec 2025, Ahmadi et al., 2022). The canonical formulation is:

$\min_x Z(x) = \sum_{i\in I} p_i x_i$

subject to: $$\sum_{i\in I} c_{i,n} x_i \ge R_n, \quad \forall n\in N$$

$\sum_{i\in I} p_i x_i \le B_w$

$x_i = 0, \quad \forall i \in I_{\text{excluded}}$

$x_i \ge 0$

Inverse optimization approaches extend this by jointly learning utility functions $c_k$ for user clusters and producing representative plans $x^*_k$ that satisfy personalized constraints, leveraging empirical intake data and projection of infeasible observations (Ahmadi et al., 2022).

Food substitution graphs $G=(V,E)$ allow efficient replacement of constrained or costly items via shortest path and cost-penalty minimization algorithms:

$$w_{ij} = \alpha \frac{p_j - p_i}{p_i} + \beta \sum_{n\in N} |c_{i,n} - c_{j,n}|$$

Dynamic adaptation is a core feature: when price updates $\Delta p_i$ introduce shocks beyond threshold $\tau \approx 10\%$, meal plans are re-solved in under 2 seconds on 300+ items using industry-standard LP solvers (Syed et al., 24 Dec 2025).

3. Knowledge Representation and Multi-Agent Reasoning

AgentDiet architectures leverage knowledge graphs and ontology-driven knowledge bases for efficient routing and evidence retrieval (Shi et al., 10 Oct 2025, Singh et al., 2021, Xu, 8 Jan 2026, Fadhil, 2018). The NG-Router paradigm formalizes multi-agent collaboration for nutrition QA as a supervised GNN-guided routing over heterogeneous knowledge graphs:

$G = (V, E), \quad V = V_Q \cup V_A \cup V_E$

Routing distribution $r_i = p_\theta(a_i|q,G)$ is learned via message passing over agent, query, and entity nodes, while gradient-based subgraph retrieval mechanisms prune irrelevant or noisy context using salience scores $$\alpha_i' = \sigma(\|\nabla_{h_i^{(L)}} L_{\mathrm{route}}\|_2)$$ (Shi et al., 10 Oct 2025). This results in highly scalable, interpretable, and modular systems in which new agent nodes or dietary evidence can be added with minimal retraining or orchestration code.

Conversational AgentDiet frameworks for dietary supplements employ structured knowledge bases (iDISK), fine-grained NLU models (CNN/CRF), and rule-based dialogue policies for contextually accurate and efficient response generation (Singh et al., 2021). Template-driven natural language generation ensures factual consistency.

4. Inference-Time Trajectory Efficiency Mechanisms

AgentDiet also refers to a trajectory reduction module for LLM-based agents, targeting input token and computational efficiency in multi-step tool-invocation systems (Xiao et al., 28 Sep 2025). The trajectory $T = [T[0], ..., T[s]]$ contains both agent and tool messages, and is pruned via reflection (a secondary LLM) to remove useless, redundant, and expired information:

• Useless: non-contributory tokens (e.g., cache listings)
• Redundant: repeated arguments or outputs
• Expired: previously relevant but currently obsolete context

Empirically, AgentDiet reduces input tokens by 39.9–59.7%, agent computational cost by 21.1–35.9%, with no statistically significant change in task success rate. Plug-and-play deployment, support for a broad agent spectrum, and cache-optimized sliding-window editing are distinguishing features (Xiao et al., 28 Sep 2025).

5. Personalized Nutrition Management and Behavioral Coaching

Recent AgentDiet systems employ multi-agent controllers with closed-loop feedback for meal-level personalization, image-based nutrient estimation, and dynamically updated intake budgets (Xu, 8 Jan 2026, Yang et al., 2024). A typical cycle includes:

1. User submits meal image/description.
2. Vision Agent estimates portions/ingredients and nutrient vector $N_{\text{meal}}$ using reference objects for scale calibration.
3. State Management Agent updates cumulative intake $A_t$ and remaining budget $B_t$.
4. Dialogue/Recommendation Agent proposes next meals or solicits clarification.
5. Controller Agent ensures optimal task sequence and policy.

Nutrient estimation achieves high coverage ($\approx 0.96$ core macros) and MAE errors of 58.9 kcal (energy), 6.8 g (protein), 225 mg (sodium). End-to-end latency averages 65.4 s, and directional agreement with ground truth is significantly better than random baselines (Xu, 8 Jan 2026).

Behavioral science-informed workflows split coaching into Barrier Identification Agent (BIA) and Strategy Execution Agent (SEA), mapping individual barriers $B = \{b_1,...,b_{28}\}$ to evidence-based tactics via mappings $f_1: B \to \mathcal{P}(S)$ and $f_2: S \to \mathcal{P}(T)$, achieving barrier identification accuracy of 0.93 and tactic personalization Likert scores $\geq 4.38$ (Yang et al., 2024).

6. Evaluation, Scalability, and Generalization

Robust empirical evaluation is documented across various AgentDiet implementations:

• Household Cost/Nutrient Adequacy: Saudi case study shows 17% weekly cost reduction and $\geq$95% adequacy under $\pm$(20–30)% price shocks (Syed et al., 24 Dec 2025).
• Conversational QA: Supplement agent achieves 76.2% succ@3+ and inter-annotator agreement $\kappa=0.82$ (Singh et al., 2021).
• LLM Trajectory Efficiency: Input tokens down 39.9–59.7%, costs down 21.1–35.9%, pass rates stable (Xiao et al., 28 Sep 2025).
• Image-Based Logging: MAC nutrient coverage 0.96, mean portion error 17.4% (Xu, 8 Jan 2026).
• Behavioral Coaching: Personalization and actionability Likerts $>$4.2, barrier identification accuracy $>$0.90 (Yang et al., 2024).
• Inverse Optimization: Cluster-based recommendations strictly enforce nutrition constraints, outperforming standard clustering in adherence and health alignment (Ahmadi et al., 2022).

AgentDiet’s modular design supports region-specific extension (nutrient tables, price feeds), health condition-specific modules (renal, FODMAP), and household scaling (multigenerational, income segments). Technical scalability relies on containers, autoscaling, and horizontal scaling for thousands of users (Syed et al., 24 Dec 2025).

7. Ethical, Behavioral, and Practical Considerations

AgentDiet frameworks incorporate established behavioral change theories (e.g., Schutz von Thun, PSD Framework, BCT Taxonomy), tailoring user engagement through micro-goals, personalized reminders, and social sharing (Fadhil, 2018). Ontology-driven data minimization, privacy (TLS+AES-256 encryption), anonymization, and explicit user consent are recognized pillars.

Implementation guidelines emphasize template-based NLG, fail-safe fallback responses, iterative prototyping, and Bayesian/Multi-Armed Bandit adaptation for feedback-driven personalization. Proposed future directions include multimodal data fusion (CGM, images), EHR integration for clinical adoption, and adaptive inference scheduling in trajectory efficiency modules (Xu, 8 Jan 2026, Xiao et al., 28 Sep 2025, Yang et al., 2024, Fadhil, 2018).

In sum, AgentDiet comprises a comprehensive, empirically validated family of agentic architectures for personalized, efficient, and adaptive nutrition management across heterogeneous domains, with modularity and scalability at its core (Syed et al., 24 Dec 2025, Shi et al., 10 Oct 2025, Ahmadi et al., 2022, Xu, 8 Jan 2026, Singh et al., 2021, Yang et al., 2024, Xiao et al., 28 Sep 2025, Fadhil, 2018).