A Comprehensive Survey of Agentic Time Series Systems: Architecture, Reliability, and Research Frontiers

1 Introduction

Financial markets, medical monitoring, transportation networks, industrial equipment, energy systems and climate environments all expose their state as time series. Real‑world tasks require more than next‑step prediction: they must assess data quality, detect anomalies, link external events, select models and tools, quantify uncertainty, explain results, and, when needed, take actions with long‑term consequences.

The paper groups the development of time‑series intelligence into four intertwined routes:

Benchmark Evaluation : expands from prediction error to reasoning correctness, tool reliability and decision safety.

Foundational Models : learns universal time representations and cross‑domain forecasting ability.

LLM4TS : connects numeric, linguistic and explicit temporal reasoning spaces.

Time‑Series Agents : closes the loop with perception, reasoning, tool use, action, feedback and state update.

2 Preliminaries

Four Time‑Series System Paradigms

The paper first distinguishes four easily confused paradigms.

Pre‑trained Backbone Models : directly learn time representations or predictive distributions from raw series (e.g., Chronos, MOMENT, TimesFM). Strong at generic forecasting but lack tool, memory and action loops.

LLM Translators : map series to text, prompts, images or symbolic forms so that LLMs can process temporal information (e.g., Time‑LLM, TEST, ChatTS). Improves interaction and semantic alignment yet remains a single‑shot I/O process.

LLM Reasoners : perform explicit inference over trends, cycles, anomalies, similarity and causality, possibly using chain‑of‑thought, self‑critique or reinforcement learning. Provides explanations but typically misses persistent memory, tool‑driven action and feedback adaptation.

Time‑Series Agents : maintain a closed‑loop policy—observe temporal evidence, select tools or actions, receive environment feedback, update state, and decide the next step. Higher‑order capabilities such as explicit reasoning, memory/knowledge, and world modeling can be added.

Figure 3: Four paradigms of LLM‑based time‑series systems. The key distinction is whether the system forms a closed loop around evidence, action, feedback and state update.

Why Time‑Series Agents Differ

Non‑stationarity : requires detecting regime or operating‑condition shifts and promptly updating memory and models.

Delayed Feedback : demands retaining full trajectories to solve long‑term credit‑assignment.

High‑Cost Actions : imposes risk budgeting, confidence thresholds and human approval.

Uncertainty Accumulation : calls for calibration, replanning, refusal and fallback strategies.

Multi‑scale, Multi‑modal Evidence : needs alignment of numeric, event, log, chart and document modalities.

Statistical Verifiability : mandates that trend, lag and causal claims be testable.

The paper defines a time‑series agent as a closed‑loop system existing within a dynamic temporal environment. Perception and planning form the minimal core; explicit reasoning, memory/knowledge, and world modeling represent progressively enhanced capabilities. Reliability is a vertical constraint across all layers.

Figure 4: Five‑layer compositional architecture of a closed‑loop time‑series agent. Reliability and trustworthiness constrain evidence, reasoning, action, memory and simulation throughout the run.

3 Time Series Perception

The perception layer does not merely ask “how to encode the series” but asks “what temporal evidence is needed and in what form”. Five perception categories are defined:

Raw Numeric Perception : retains raw values, timestamps, channels and local windows, emphasizing fidelity and tool compatibility. Suited for precise statistics and forecasting but suffers from context length pressure and possible scale loss when textified.

Diagnostic‑Tool Perception : transforms series via decomposition, ACF/PACF, spectral analysis, change‑point detection, missing‑pattern and anomaly detection into checkable evidence. Benefits reproducibility and verification; risk lies in possible tool mis‑selection.

Symbolic Perception : compresses series into trends, cycles, events, attributes, prototypes or textual summaries for easier LLM consumption. Bottleneck is translation error—mislabeling normal fluctuations as anomalies corrupts downstream reasoning.

Structural Perception : explicitly represents variable relationships, temporal hierarchies, regimes, graph structures, topic segments and repeated patterns, enabling the agent to see cross‑channel dependencies and multi‑scale structure.

Multimodal Perception : aligns series with charts, tables, news, logs, weather, electronic health records, etc., preserving timestamps, sources and alignment relations.

Figure 5: Classification of time‑series perception. Different interfaces expose distinct types of actionable evidence.

The perception layer should output a structured evidence state rather than an opaque vector, recording source, time window, processing method, confidence and supporting material for downstream auditing.

4 Time Series Reasoning

The reasoning layer converts structured evidence into judgments about dynamic patterns, numeric relations, anomalies, causal hypotheses, uncertainty and future states.

Pattern & Structure Reasoning

Identifies trends, seasonality, cycles, spikes, regime switches and cross‑variable relations, distinguishing superficial similarity from structural similarity.

Numeric & Statistical Reasoning

Computes statistics, thresholds, correlations, frequencies and forecast metrics. High‑trust systems delegate precise calculations to executable tools rather than relying on LLM intuition.

Causal & Compositional Reasoning

Analyzes lag effects, external events, variable dependencies and candidate causal paths. Emphasizes that temporal precedence and correlation do not automatically prove causality; testable hypotheses and counterfactual evidence are required.

Reflective & Metacognitive Reasoning

Checks whether conclusions align with raw series and tool outputs, evaluates uncertainty, and retries, supplements observations, or refuses when evidence is insufficient.

Slow‑Thinking & Reinforcement Reasoning

Decomposes complex problems into multi‑step strategies, teaching the model when to call tools, which data segment to use, and how to verify intermediate conclusions. Reinforcement learning can optimise decision trajectories, but rewards must cover correctness, cost and safety.

Figure 6: The reasoning layer transforms multi‑source temporal evidence into hypotheses, explanations, uncertainty estimates and verifiable bases, feeding the decision input after reflection and tool verification.

The paper argues that trustworthy temporal reasoning should shift from evaluating only the final answer to validating every intermediate step, binding each claim to concrete windows, statistics or tool outputs.

5 Planning and Action

The planning‑action layer distinguishes agents from passive reasoners: the system must decide what to do next, not just what happened.

Workflow Planning

Decomposes open‑ended tasks into stages such as data diagnosis, preprocessing, model selection, validation and reporting, and revises the workflow based on feedback. Systems like TimeSeriesScientist and TimeCopilot make the analyst’s work explicit.

Tool Routing & Evidence Acquisition

The agent selects appropriate tools for decomposition, anomaly detection, relevance analysis, forecasting, retrieval or simulation, passing correct windows, frequencies, parameters and variables.

Model, Data & Code Orchestration

The agent can filter auxiliary series, generate code, run models, compare candidate solutions and manage experiment artifacts, producing reproducible analysis pipelines rather than mere predictions.

Multi‑Agent Collaboration

Assigns data analysis, forecasting, review, risk control and reporting to distinct roles, reducing responsibility conflicts but introducing communication cost, error propagation and consistency challenges.

External Decision Execution

In finance, traffic, energy and industry, agents may execute trades, control actions, raise alerts or allocate resources, subject to confidence, risk, permission and human‑in‑the‑loop constraints.

Figure 7: Planning‑action closed loop. The agent combines goals, evidence, memory and trajectory to select workflows, data ops, models, code, retrieval, communication or external intervention, updating state with environment feedback.

6 Memory and Knowledge

Agents need structured, selectable, updatable and forgettable experience rather than raw dialogue logs.

Context Memory : stores current session evidence, tool results and decision state.

Scenario Memory : retains historical cases and full task trajectories for similar‑scenario retrieval.

Temporal Memory : records regimes, seasonal patterns, long‑term dependencies and multi‑scale changes.

Knowledge Memory : keeps domain rules, causal structures, tool manuals and constraints.

Program Memory : saves successful workflows, strategies and reflective experiences.

Failure & Confidence Memory : logs error patterns, calibration signals and unreliable tools.

Figure 8: Six categories of memory and their interaction with perception, reasoning, planning, world modeling and verification. Memory undergoes formation, retrieval and evolution.

Current systems mostly rely on context memory or simple case retrieval. Open challenges include handling conflicts between old and new regimes, preventing stale or poisoned memories from influencing decisions, evaluating cross‑session benefits, and determining whether programmatic experience can transfer across domains.

7 Time Series World Models

Traditional forecasting answers “what will happen next”. World models ask “why does the environment evolve this way, and what would happen under changed conditions”.

The paper summarises four routes:

Environment Understanding : builds structured states linking variables, events, lags and domain constraints, providing a basis for relational reasoning. Inferred dependencies are not guaranteed to be true causal mechanisms.

Temporal Simulation : generates probabilistic future trajectories from history and context. Works like Chronicle treat prediction as forward rolling of world state; BRIDGE controls sequence generation via language and semantic prototypes.

Counterfactual Simulation : compares baseline futures with alternatives under hypothetical changes (policy, weather, market events). Validation is difficult because alternative futures are unobservable.

Executable Deployment : packages domain data, simulators and verification tools as environments callable by agents. AgriWorld demonstrates an agricultural protocol, but building such environments requires extensive domain engineering.

World modeling for time series remains early; progress may come from domain‑specific simulators rather than a universal universal model.

8 Reliability and Trustworthiness

Errors propagate in a closed loop: noise → wrong prediction → false explanation → erroneous tool use or dangerous action → polluted memory → reinforced future errors.

Figure 9: Failure propagation in agentic time‑series systems. Evidence noise, mis‑prediction, hallucination, faulty tools, unsafe actions and memory poisoning can form cycles; safety checks, controls and audit trails are needed to break them.

The paper proposes a multi‑layer reliability stack:

Prediction Quality : accuracy, calibration, long‑term stability, data leakage control.

Robustness : regime change, extreme events, sensor failure and recovery.

Reasoning Correctness : numeric validity, temporal ordering, causal discipline.

Tool Reliability : routing, parameters, execution recovery, result consistency.

Hallucination & Evidence Grounding : every fact must be backed by raw series, events or tool output.

Safety & Protection : prompt injection, tool attacks, data and memory poisoning.

Decision Safety : risk budgeting, constraints, refusal and escalation mechanisms.

Auditability : timestamps, prompts, tool versions, random seeds and reproducible artifacts.

The core principle is that reliability must be a contract between layers, not a post‑hoc check at the final output.

9 Benchmarks and Evaluation Protocols

Evaluation evolves through four stages:

Pure Sequence Baselines : predict, classify, interpolate, detect anomalies; measure numeric fidelity, generalisation and calibration.

Heterogeneous Information‑Enhanced : augment series with news, events, reports, images or metadata; still output numbers or labels, testing whether external info improves forecasts.

Understanding & Reasoning : convert outputs to QA, explanations and reasoning traces; assess structure recognition, multimodal alignment, anomaly explanation and causal inference, though still largely static single‑turn.

System‑Level Agent Evaluation : require code generation, ML pipeline construction, active evidence retrieval and tool use. Benchmarks like TimeSeriesGym, Dr‑CiK, TSAIA begin assessing full execution trajectories.

Current benchmarks mainly cover perception and reasoning; planning‑action is nascent; memory and world models lack standard protocols. Future evaluation should involve multi‑session, interactive environments that test long‑term memory, adaptive strategy, delayed feedback and counterfactual simulation.

10 Applications

The survey lists seven application domains:

General Analytics : open‑ended data diagnosis, model selection, code execution and transparent reporting.

Finance & Trading : delayed feedback, regime shifts, risk budgeting, back‑test consistency and reproducible execution.

Transportation & Urban Systems : connect forecasts to spatial reasoning, simulator configuration and control policies.

Weather, Energy & Buildings : require multi‑scale physical evidence, actionable explanations, optimisation and constraint control.

Healthcare : demand conservative decisions, guideline knowledge, confidence calibration, refusal and clinician review.

Industrial IoT & Observability : extend anomaly detection to logs, metrics and tracing for root‑cause diagnosis.

Agriculture, Automation, Retail & Supply Chain : focus on how forecasts feed simulation, validation, inventory planning and operational control.

These scenarios illustrate that lower prediction error does not guarantee better decisions; systems must jointly evaluate action cost, risk, latency and traceability.

11 Positions and Frontiers

The paper proposes six core research frontiers:

Verifiable Temporal Reasoning : shift from final‑answer correctness to step‑level evidence grounding; each trend, lag or causal claim must be checkable.

End‑to‑End Agent Training : move from handcrafted prompt‑tool pipelines to joint optimisation of perception, planning, memory and action, addressing credit assignment under delayed feedback.

Memory & Lifelong Adaptation : build proactive cross‑session memory with clear retain, update, forget and conflict‑resolution policies, distinguishing genuine transfer from data leakage.

Time‑Series World Models : evolve from point forecasts to generative, counterfactual‑queryable, mechanism‑verifiable environment models.

Decision‑Centric Evaluation : jointly measure prediction, confidence and action quality under distribution shift, partial observability and operational cost.

Standardised Temporal Tool Protocols : unify data frequency, units, missing‑value handling, window semantics, model versioning, confidence and provenance to enable interoperable, verifiable and reproducible tools.

The overarching stance is that time‑series agents should be regarded as "time decision systems". The true frontier lies not in larger predictive models but in integrating the five‑layer capabilities under reliability constraints to form learnable, auditable and sustainably improving closed loops.

12 Conclusion

The survey delineates clear boundaries for agentic time‑series systems: foundational models provide universal representations, translators bridge numeric and linguistic spaces, reasoners interpret temporal evidence, and agents close the loop through tools, actions, feedback and state updates.

The five‑layer architecture offers a unified coordinate system for existing work. Perception decides what the system sees, reasoning decides what conclusions can be drawn, planning and action decide the next step, memory decides how experience is accumulated, and world modeling decides how to compare possible futures; reliability constrains each layer’s trustworthiness.

Future progress is not about smoother textual output but about reliable analysis, decision‑making, learning and action in dynamic uncertain environments. When evidence is traceable, reasoning verifiable, tools reproducible, actions constrained, and memory auditable, time‑series agents can move from research prototypes to high‑risk real deployments.