SimAgents: Heterogeneous Simulation Agents

• SimAgents are simulation frameworks that use explicit agent abstractions to model dynamic decision processes across various domains.
• They are applied in autonomous driving, economic modeling, smart grids, and LLM-mediated workflows to tackle domain-specific simulation challenges.
• Research in SimAgents addresses scalability, realism, and evaluation issues via distributed computing, graph-based interactions, and innovative multi-agent modeling.

Searching arXiv for papers explicitly using the term “SimAgents” and closely related simulator frameworks. SimAgents denotes a heterogeneous research area rather than a single canonical system. In recent literature, the term is used for simulated actors in closed-loop trajectory rollouts, for heterogeneous agents in economic, infrastructure, and smart-grid environments, and for LLM-based multi-agent systems that configure simulations, validate parameters, interpret outputs, or generate dynamic scenes (Zhao et al., 2024, Dwarakanath et al., 2024, Zhang et al., 11 Jul 2025, Liu et al., 2 Jul 2026). Across these uses, the common substrate is an explicit model of state, action, transition, and feedback, but the operational meaning of an “agent” varies substantially: it may be a vehicle, a household, a market participant, a graph-local controller, a cohort-level abstraction, or a specialized software role.

1. Terminological scope and research uses

Across the cited literature, “SimAgents” is used in non-identical senses. In autonomous driving, it is tied to the Waymo SimAgents Challenge, whose objective is to generate realistic closed-loop trajectories for all road agents over an 8 s horizon, given 2 s of real driving history (Zhao et al., 2024). In economics and infrastructure, it refers to heterogeneous agents such as Households, Firms, Central Bank and Government agents, or to modular agent classes in agent-based market and smart-grid simulators (Dwarakanath et al., 2024, Trimborn et al., 2018, Chassin et al., 2014). In LLM-centric work, it can denote specialized software agents that extract cosmological parameters from papers, rewrite negotiation strategies, or coordinate staged 4D scene construction (Zhang et al., 11 Jul 2025, Mangla et al., 5 Oct 2025, Liu et al., 2 Jul 2026).

Usage	Representative systems	Defining feature
Closed-loop road-agent simulation	KiGRAS, MPS	8 s rollout, physical feasibility, interaction realism
Economic and infrastructure simulation	ABIDES-Economist, SABCEMM, GridLAB-D, MAXE, SimCity	Heterogeneous agents, markets, event-driven or Gym-style execution
LLM-mediated simulation tooling	SimAgent, Simulation Agent, NegotiationGym, SimWorlds	Tool calling, prompt rewriting, planner-coder-reviewer workflows
Scalable or abstracted multi-agent engines	TeraAgent, GAMMS, GA-S$^3$	Distributed execution, graph locality, or group-agent compression

This breadth suggests that SimAgents is best understood as a family resemblance category. What unifies the family is not domain or implementation language, but the use of explicit agent abstractions to represent local decision processes inside a larger dynamical system.

2. State, action, and transition formulations

The mathematical formulation of SimAgents varies sharply by domain. In ABIDES-Economist, the environment is cast as a finite-horizon Partially Observable Markov Game,

$$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$

with agents selecting actions from observations and seeking to maximize discounted reward over the horizon (Dwarakanath et al., 2024). In GAMMS, the global state is

$s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$

and transitions are deterministic,

$$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$

implemented by graph-constrained motion plus user-defined rule functions (Patil et al., 4 Feb 2026).

In trajectory simulation, KiGRAS replaces direct state-space autoregression with action-space modeling. It introduces control actions $U_t=(A_t,Y_t)$, a deterministic kinematic map $S_{t+1}=K(S_t,U_t)$, and a training objective

$$\max_\theta \prod_{t=0}^{T-1} P_\theta(U_t \mid S_{\le t}, W_{\le t}) \quad\text{subject to}\quad S_{t+1}=K(S_t,U_t).$$

The future joint over states and actions is factorized as

$$P(S_{1:T}, U_{0:T-1} \mid history) = \prod_{t=0}^{T-1} P(U_t \mid S_{\le t}, W_{\le t}) \cdot \delta(S_{t+1} - K(S_t, U_t)),$$

which hard-enforces physical reachability at every step (Zhao et al., 2024).

Event-driven simulators adopt a different formalism. In MAXE, a message is a tuple $m=(s,r,\tau,\delta,c)$ with delivery time $t_{\text{deliver}}=\tau+\delta$, and agent updates occur on message delivery through a global priority queue (Belcak et al., 2020). This suggests that SimAgents research is not organized around one dominant simulation kernel; instead, discrete-event scheduling, Markov-game formulations, deterministic graph transitions, and receding-horizon kinematic rollout coexist as equally standard choices.

3. Mobility, traffic, and closed-loop realism

The most technically explicit use of SimAgents appears in road-user simulation. The Waymo SimAgents Challenge evaluates closed-loop simulation on WOMD v1.2, with 103 K real-world scenes, 2 Hz sampling, and metrics Kin, Inter, Map, Realism Meta-Score, and minADE (Zhao et al., 2024). KiGRAS addresses the task by discretizing longitudinal acceleration and yaw rate into a 63 $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$0 63 action vocabulary, yielding 3 969 possible actions. Using the CTRA formulation,

$$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$1

$$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$2

all generated states are guaranteed physically reachable. On the SimAgents metrics comparison, KiGRAS-0.7M reports Real. 0.7597, Kin. 0.4691, Inter. 0.8064, Map. 0.8658, and minADE 1.4383, attaining the top Realism score with 0.7 M parameters (Zhao et al., 2024).

A complementary line is Model Predictive Simulation. MPS combines an MTR++ transformer proposal model with a probabilistic graphical model containing factors for smooth trajectories, avoidance of collisions with static obstacles and other moving agents, and approximate MAP inference via Gauss-Newton (Lou et al., 2024). It samples $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$3 trajectories for each of the $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$4 agents for the next $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$5 time steps, with $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$6, returns only the first predicted step, and replans in receding-horizon fashion. On the WOSAC leaderboard excerpt reported in the paper, MPS reaches Realism 0.7416, COLLISION 0.9629, OFFROAD 0.9366, and minADE 1.4841, and an ablation against MTR+RAND shows large gains in collision and offroad metrics (Lou et al., 2024).

Beyond vehicle-level rollout, recent work extends SimAgents to city-scale mobility choice. GTA constructs a census-grounded synthetic population from the “Mobility in Germany 2017” survey using the “Truncate, Replicate, Sample” method, generates natural-language personas with an LLM, produces activity schedules with structured few-shot prompting and Chain-of-Thought scaffolding, and selects modes from route candidates produced by SUMO or OTP (Lämmer et al., 23 Jan 2026). In Berlin-scale experiments, GTA reports overall mode shares of walk 27.2 %, bike 20.6 %, car 31.5 %, and PT 20.7 %, with $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$7, while also exhibiting systematic biases in trip length and mode preference (Lämmer et al., 23 Jan 2026).

Earlier traffic simulators used more explicit cognitive architecture. “A multiagent urban traffic simulation” builds a topological transport graph from GIS data and instantiates immobile Transporter agents on road elements and Mobile agents for vehicles. Mobiles maintain beliefs, desires, and intentions, revise edge weights under congestion, and switch among rerouting tactics such as least-used roads or flocking follow (Tranouez et al., 2012). A plausible implication is that contemporary traffic SimAgents have shifted from handcrafted BDI-style adaptation toward learned or hybrid generative-control formulations, while retaining the same core concern with dynamic replanning under interaction.

4. Economic, market, and infrastructure SimAgents

Economic simulation has developed its own distinct SimAgents tradition. SABCEMM formalizes an abstract ABCEM model in terms of agents, an excess-demand calculator,

$$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$8

and a price calculator that may be a rational market solving $$\Gamma = \bigl\langle \mathcal{N}, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, \mathbb{T}, \{\mathbb{O}_i\}, \{R_i\}, \{\beta_i\}, H\bigr\rangle,$$9 or an irrational market following a stochastic update rule (Trimborn et al., 2018). Its software architecture centers on three abstract C++ classes—Agent, ExcessDemandCalculator, and PriceCalculator—wired together by XML configuration, with computational complexity strictly $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$0 for all non-networked models and support for several million agents (Trimborn et al., 2018).

ABIDES-Economist generalizes the economic setting to heterogeneous Households, heterogeneous Firms, Central Bank and Government agents, all orchestrated by an interactive discrete-event kernel with timestamped message passing (Dwarakanath et al., 2024). Firms experience stochastic technology shocks through a log-AR(1) process, households optimize utility over consumption, labor, and savings, and learning agents are integrated through an OpenAI Gym–style API with reset() and step(a) methods. The framework is explicitly designed so that agents can be rule-based or reinforcement-learned in the same simulator (Dwarakanath et al., 2024).

MAXE represents another design point: a C++ core with a message-driven event queue, pybind11 Python bindings, XML-based configuration, and exchange agents maintaining limit-order books with balanced binary search trees (Belcak et al., 2020). Message scheduling and delivery cost $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$1 for $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$2 pending messages, and order-book operations cost $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$3 for $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$4 outstanding orders. The paper reports that MAXE scales roughly linearly in population size, uses approximately 100 MiB for 100 022 agents when implemented in pure C++, and enables both financial-market microstructure studies and non-financial routing experiments (Belcak et al., 2020).

Infrastructure-oriented SimAgents are exemplified by GridLAB-D, where electrical network agents, generation and storage agents, load agents, building agents, market agents, and optional communication agents participate in a rank-ordered, event-driven solver (Chassin et al., 2014). Each agent maintains a next event time $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$5 and a valid-to horizon $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$6, and the solver advances to the minimum reported event time. GridLAB-D thereby couples power flow, thermal dynamics, market clearing, and communication effects inside one agent-based simulation environment (Chassin et al., 2014).

Recent LLM-based macroeconomic simulators continue this line in a different idiom. SimCity uses four LLM-powered roles—households, firms, a government, and a central bank—alongside a VLM for geographic placement of new firms in a city map (Feng et al., 1 Oct 2025). The environment includes a frictional labor market, a heterogeneous goods market with 44 differentiated goods, and a core financial market. The framework reproduces stylized relationships including the Phillips Curve, Okun’s Law, the Beveridge Curve, price elasticity patterns, and Engel’s Law, with robustness across random seeds (Feng et al., 1 Oct 2025). This suggests that economic SimAgents research now spans a continuum from classical object-oriented ABMs to language-mediated macro systems.

5. LLM-mediated simulation agents and software roles

A major recent expansion of SimAgents research treats agents not as entities inside the simulated world, but as software roles around the simulation process itself. The “simulation agent” framework defines a Simulation Agent as an AI-powered software component between a user and a standalone simulation model, with a parser $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$7, a tool scheduler $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$8, and a generator $s_t = (G, \{\ell_i(t)\}, \{m_i(t)\}),$9 operating over a simulation $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$0 (Kleiman et al., 19 May 2025). Its tool layer exposes run_simulation(), modify_inputs(), and interpret_outputs(), with GPT-4o accessed via LangChain and no further fine-tuning reported. The reported evaluation is preliminary but includes qualitative findings that non-technical users could configure scenarios in plain English and that simulation iteration time was reduced by 30–50% as the agent managed batch runs in parallel (Kleiman et al., 19 May 2025).

The cosmology system "Bridging Literature and the Universe Via A Multi-Agent LLM System" uses the name SimAgent for a more specialized pipeline: a Physics Agent extracts parameters from papers, a Software Agent validates and formats them against MP-GADGET requirements, and an Analysis Code Writer produces preliminary analysis code from outputs (Zhang et al., 11 Jul 2025). The benchmark contains $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$1 cosmological simulations manually annotated into .genic and .gadget parameter sets. Against Chain-of-Thought and Exchange-of-Thought baselines, SimAgent reports Micro-F1 98.67%, Precision 97.80%, Recall 99.55%, with average errors per case of 0.46 value, 0.02 type, and 0.30 hallucination (Zhang et al., 11 Jul 2025).

NegotiationGym shows a different software-role interpretation. Each agent implements compute_utility(E) and optionally learn_from_feedback(E), with self-optimization typically realized by collecting the last $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$2 episodes, prompting an LLM as a negotiation coach, and rewriting the system prompt for future rounds (Mangla et al., 5 Oct 2025). In a buyer–seller study over 20 episodes and four optimization modes, cumulative average private utilities improve whenever an agent is being coached, and both-reflect mode moves outcomes closer to the Pareto frontier while minimizing no-deal cases, although no formal proof of convergence is given (Mangla et al., 5 Oct 2025).

SimWorlds extends the same paradigm to dynamic 3D scene generation. It uses Planner, Coder, Reviewer, and Final Reviewer roles plus an orchestrator and deterministic verifier, arranged in a fixed stage sequence aligned with Blender’s dependency graph (Liu et al., 2 Jul 2026). The verifier checks structural, geometric, and plan-versus-state rules, while runtime-state inspection tools expose modifier stacks, baked state, f-curves, and motion-sheet previews. On 4DBuildBench, SimWorlds reports overall MPR 0.87 vs 0.67, SPR 0.89 vs 0.70, and VLM 0.82 vs 0.78 against VIGA, indicating that staged verification matters especially for mechanism correctness (Liu et al., 2 Jul 2026).

6. Scalability, abstraction, and systems engineering

SimAgents systems pursue scale through markedly different engineering strategies. TeraAgent is a distributed agent-based simulation engine in which MPI ranks hold local agent containers, uniform neighbor-search grids, and partitioning grids, while halo exchange and agent migration handle cross-partition interactions (Breitwieser et al., 28 Sep 2025). Its tailored serialization performs one pass over an agent subtree and one linear scan on deserialization, both $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$3 in serialized size, and delta encoding reduces raw serialized size by up to 3.5$$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$4 before LZ4 compression. The reported scale reaches 102.4 billion agents on 128 nodes at 7.08 s per iteration and 501.5 billion agents on 438 nodes at 147 s per iteration, using 92 TB memory (Breitwieser et al., 28 Sep 2025).

GAMMS emphasizes lightweight graph-structured simulation rather than extreme distributed count. It represents the world as a graph $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$5, attaches agents to nodes or edges, exposes all modules through a single Context object, and separates GraphEngine, AgentEngine, SensorEngine, adversarial rules, and visualization (Patil et al., 4 Feb 2026). The per-timestep complexity is

$$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$6

and on graphs with $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$7, $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$8, and $$\bar T: S \times A \to S,\qquad s_{t+1} = \bar T(s_t,a_t),$$9, the framework sustains 500–1 000 steps/sec on a 4-core laptop; a published benchmark reports approximately 800 steps/s for $U_t=(A_t,Y_t)$0, $U_t=(A_t,Y_t)$1, and $U_t=(A_t,Y_t)$2 in headless mode (Patil et al., 4 Feb 2026).

GA-S$U_t=(A_t,Y_t)$3 attacks scale by replacing individual agents with Group Agents. Instead of simulating all users, it maintains groups with shared perception, state, and memory, governed by a Markov-Network thought process and LLM-based policy (Zhang et al., 4 Jun 2025). The complexity changes from $U_t=(A_t,Y_t)$4 at the individual level to $U_t=(A_t,Y_t)$5 at the group level, with reduction ratio

$U_t=(A_t,Y_t)$6

On the SNB benchmark, GA-S$U_t=(A_t,Y_t)$7 reports MAPE 16.48%, DTW-mean $U_t=(A_t,Y_t)$8, and Z-score $U_t=(A_t,Y_t)$9 across five repeated runs (Zhang et al., 4 Jun 2025).

These systems illustrate three distinct scaling doctrines: distribute the same agent model across many nodes, restrict interaction to graph-local neighborhoods, or compress many individuals into cohort-level surrogates. No single doctrine dominates across domains.

7. Evaluation regimes, recurrent limitations, and misconceptions

Evaluation in SimAgents research is strongly domain-specific. Autonomous-driving systems emphasize closed-loop realism, with Kin, Inter, Map, Realism Meta-Score, minADE, collision, and offroad metrics (Zhao et al., 2024, Lou et al., 2024). Social-network group simulation uses paired t-test, MAPE, Dynamic Time Warping, and Z-score (Zhang et al., 4 Jun 2025). SimWorlds separates visual fidelity from physical consistency through Mechanism Pass Rate, Structural Pass Rate, and an itemized VLM judge (Liu et al., 2 Jul 2026). Cosmology-oriented SimAgent evaluates precision, recall, $S_{t+1}=K(S_t,U_t)$0, and error breakdowns into value error, type error, and hallucination (Zhang et al., 11 Jul 2025). GTA compares modal split, trip-length and duration distributions, traffic counts, and an interregional benchmark, reporting aggregate RMSE $S_{t+1}=K(S_t,U_t)$1 and correlation $S_{t+1}=K(S_t,U_t)$2 for morning-peak shape in one district (Lämmer et al., 23 Jan 2026).

A recurrent misconception is that plausible surface behavior is sufficient evidence of simulation correctness. SimWorlds directly contests this by showing that visual judges cannot detect mechanism cheating and by introducing engine-state audit rules that inspect solver state rather than rendered output alone (Liu et al., 2 Jul 2026). A related misconception is that larger models are necessarily better simulators: KiGRAS reaches the top Realism score on the Waymo leaderboard with 0.7 M parameters, and its core claim is that action-space factorization removes redundant state-space variation rather than enlarging model scale (Zhao et al., 2024).

The limitations reported across the literature are equally heterogeneous. KiGRAS notes that discrete action bins may limit resolution in highly dynamic maneuvers, that MPC-based inverse kinematics is relatively expensive, and that non-vehicular agents may require alternate control definitions (Zhao et al., 2024). GTA reports a “role-model paradox”: agents overuse cycling, underuse short walks, and rarely travel at night (Lämmer et al., 23 Jan 2026). SimAgent for cosmology retains a nonzero hallucination rate and explicitly requires trust-and-verify (Zhang et al., 11 Jul 2025). NegotiationGym reports empirical improvement curves but no formal convergence proof (Mangla et al., 5 Oct 2025). GridLAB-D notes no global-convexity guarantee and possible inefficiency when many agents synchronize on very small $S_{t+1}=K(S_t,U_t)$3 (Chassin et al., 2014). SimCity states that LLM agents occasionally produce outlier actions requiring heuristic filters (Feng et al., 1 Oct 2025).

Taken together, these limitations indicate that SimAgents is not a solved methodology but a rapidly diversifying design space. The dominant research questions now concern how to preserve causal validity and software correctness while expanding scale, heterogeneity, and agent autonomy, and how to evaluate systems whose realism may depend as much on internal consistency and intervention response as on externally plausible trajectories or dialogues.