AgentBot: Architecture and Applications

• AgentBot is a class of software agents that translate user intent into bounded, tool-using procedures through structured dialogue management.
• They span diverse architectures—from retrieval-based systems to enterprise and GUI automation—demonstrating varied applications in chemical informatics, insurance, and robotics.
• Evaluations show strong task performance and efficiency gains, yet challenges remain in context management, safety protocols, and cross-application coordination.

AgentBot denotes a class of software agents that mediate between natural-language interaction and external computational environments through dialogue management, tool invocation, and workflow control. In the cited literature, the term spans domain-specific conversational systems that query structured repositories, agent-assist interfaces grounded in enterprise data, GUI and mobile automation agents, multi-agent orchestration frameworks, and forum-mediated agents that dispatch physical robot actions. The common thread is the conversion of user intent into bounded, tool-using procedures over external state rather than unconstrained text generation alone (Sharma et al., 2021, Bradley et al., 24 Dec 2025, Yi et al., 8 May 2025, Crawford et al., 2024, Ruan et al., 3 Feb 2026).

1. Conceptual scope and historical lineages

The literature does not use a single canonical definition of AgentBot. Instead, several lineages appear. One lineage is retrieval-oriented conversational assistance. The 2017 Intelligent Personal Assistant is a retrieval-based conversational agent augmented with a knowledge navigation module and multi-modal I/O; it accepts text, voice, and Facebook input, retrieves replies from a corpus of Friends SRT files through similarity or Levenshtein matching, and routes factual “Who/What/When” queries to web scraping (Kumar et al., 2017). A later domain-specific example is Atreya, a Telegram-hosted chemistry bot that guides users through ChEMBL with rule-based AIML dialogue and ChEMBL API calls, returning identifiers, formulae, approval metadata, and molecule images (Sharma et al., 2021).

A second lineage is agent-assist in regulated or enterprise settings. Axlerod is explicitly positioned as an agent-facing, retrieval-augmented, LLM-powered assistant for independent insurance agents, with tool calls for policy lookup, policy search, and documentation search inside Safety Insurance’s environment (Bradley et al., 24 Dec 2025). HalluDetect’s AgentBot extends this enterprise orientation into consumer grievance and legal assistance by decomposing a dialogue into specialized roles—Receptionist, Paralegal, Lawyer, and Drafter—over a shared retrieval layer, with the explicit objective of reducing hallucinations rather than maximizing breadth of free-form generation (Anaokar et al., 15 Sep 2025).

A third lineage treats AgentBot as an embodied or autonomous controller. MobileAgent, EcoAgent, and AppAgent-Pro all operate on GUIs rather than only text corpora, but they differ in control philosophy. MobileAgent emphasizes human confirmation and SOP-conditioned action prediction on mobile interfaces (Ding, 2024). EcoAgent divides mobile automation across a cloud Planning Agent and edge Execution and Observation agents in a closed loop (Yi et al., 8 May 2025). AppAgent-Pro shifts from reactive instruction following toward proactive multi-app information mining and multimodal synthesis (Zhao et al., 26 Aug 2025). A related but more abstract development appears in BMW Agents and AOrchestra, where the salient unit is not a single bot but an orchestrated system of planners, executors, verifiers, memories, and dynamically created sub-agents (Crawford et al., 2024, Ruan et al., 3 Feb 2026).

This distribution suggests that “AgentBot” is best understood not as a product category with fixed boundaries, but as a family of architectures that bind language interfaces to structured action spaces. A plausible implication is that the term has moved from chat-centric assistance toward orchestration-centric automation as tool use, external memory, and verifiable execution became central design concerns.

2. Core architectural patterns

Several recurring architectural patterns organize AgentBot systems. The simplest is deterministic dialog routing over a structured backend. Atreya uses AIML for keyword-triggered intent detection, explicit prefixes such as msy, ms, mi, mc, and mg for parameter extraction, and then maps the resulting query to the ChEMBL web resource client; CairoSVG renders molecular structures and Pandas prepares filtered outputs and CSV exports (Sharma et al., 2021). This architecture minimizes ambiguity by restricting both dialogue state and query construction.

A second pattern is tool-mediated LLM orchestration in which the model chooses among explicitly exposed functions. Axlerod exemplifies this design: a Gemini 2.5 Pro model on Vertex AI receives user input plus metadata for the tools policy_detail, policy_search, and documentation_search, while LiteLLM and the Smoltalk microframework translate these capabilities into an OpenAI-compatible tool-calling interface (Bradley et al., 24 Dec 2025). In AutomationBench, the same principle is isolated as a benchmarked competency: agents must use a BM25-based Search tool over API schemas and an Execute tool with {method, url, body} to complete cross-application business workflows under policy constraints (Shepard et al., 21 Apr 2026).

A third pattern is explicit decomposition into planning, execution, observation, and verification. EcoAgent formalizes this with a cloud Planning Agent and two edge agents. Its loop is written as

$$P_0 = \text{ReACT}(Ins, S_0), \quad O_t = EA(S_t, ST_t), \quad R_t = OA(S_{t+1}, EX_t),$$

with screen compression by

$T_{t+1} = \text{PreUnderstand}(S_{t+1}),$

and failure-triggered replanning through

$$P_n = \text{Reflection}(Ins, P_{n-1}, \text{Memory}(T_1,\ldots,T_{t+1})).$$

The key mechanism is that the edge Observation Agent converts a raw screen that would cost \>1400 tokens/image through an MLLM into a 50–150 token textual summary before sending it back to the cloud planner (Yi et al., 8 May 2025).

A fourth pattern moves planning abstractions into the prompt itself. MobileAgent injects Standard Operating Procedures into the input context rather than asking the model to emit plans in the output. Its guiding relation is

$H(Y|X, Z) \leq H(Y|X),$

where $Z$ is the SOP abstraction supplied in context. In practice, this allows the model to predict concise actions such as DUAL_POINT, TYPE, SCROLL, or status signals while maintaining awareness of subtask state (Ding, 2024).

A fifth pattern is orchestration as a first-class control problem. BMW Agents decomposes the system into Coordinator, Planner, Task Queue, Executor, Agent Units, Matchers, Toolbox Refiners, Verifier, Short Memory, and Episodic Memory (Crawford et al., 2024). AOrchestra makes this even more explicit by representing any agent as the tuple

$\Phi = (I, C, T, M),$

where instruction, curated context, tools, and model are selected per subtask. The orchestrator’s action space is reduced to

$$\mathcal{A}_{AOrchestra} = \{Delegate(\Phi),\ Finish(y)\},$$

and sub-agents are created on demand rather than statically predefined (Ruan et al., 3 Feb 2026).

3. Interaction substrates and tool ecologies

AgentBot systems differ substantially in the substrate on which interaction occurs. Some are embedded in messaging applications. Atreya is deployed on Telegram through Telethon; interaction starts with /start, proceeds through button grids such as Molecule Info, Tissue Info, Similar Compound, Approved Drugs, Chat to Bot, and Exit, and combines button-driven flows with typed prefixes for structured parameter entry (Sharma et al., 2021). Others are embedded in internal enterprise web applications. Axlerod was prototyped with Open WebUI and Tampermonkey, then embedded as a chat widget inside the agent workflow so that references like “this policy” could be resolved from current-policy UI context (Bradley et al., 24 Dec 2025).

A different substrate is the GUI itself. MobileAgent predicts operations over DOM-like environment descriptions and mobile screens (Ding, 2024). EcoAgent performs action grounding and observation directly from screenshots on AndroidWorld, using image inputs without an external accessibility tree or Set-of-Mark overlay (Yi et al., 8 May 2025). AppAgent-Pro uses mobile apps such as YouTube and Amazon as external information sources and returns both textual findings and screenshots through a Streamlit-based web interface (Zhao et al., 26 Aug 2025). AgentA/B goes further by treating real webpages as experimental environments in which LLM agents search, click, filter, purchase, and stop via Selenium-controlled headless Chrome sessions (Wang et al., 13 Apr 2025).

Forum software can also become the interaction layer. AgentRob turns an online forum into the command substrate for embodied agents. Posts mentioning handles such as @quadruped or @humanoid are read through MCP-backed tools like list_posts and get_topic; commands are extracted, dispatched to robot controllers, and summarized back into the thread with metadata tags (Liu et al., 14 Feb 2026). In this setting, the forum contributes asynchronous persistence, visibility, and multi-agent coexistence rather than only dialogue.

Across these settings, tool exposure is central. Some systems rely on official client libraries, as with ChEMBL’s Python client in Atreya (Sharma et al., 2021). Others standardize tool access through schemas and middleware, as in Axlerod’s function interface (Bradley et al., 24 Dec 2025), BMW Agents’ prompt-described tools (Crawford et al., 2024), OptAgent’s 72 [MCP](https://www.emergentmind.com/topics/model-context-protocol-mcp-assisted-framework) tools for building operations (Jiang et al., 27 Jan 2026), and AOrchestra’s on-demand curation of $T$ inside each delegated tuple (Ruan et al., 3 Feb 2026). A plausible implication is that AgentBot research increasingly treats the interface contract around tools—not just the model—as the main determinant of reliability and scope.

4. Representative application domains

The breadth of the literature is most visible at the application level.

Domain	System	Implemented scope
Chemical informatics	Atreya	ChEMBL queries for drugs, targets, tissues, similarity, approvals, CSV export
Insurance agent-assist	Axlerod	Policy retrieval, billing plan, AutoPay, covered vehicles, documentation search
Consumer grievance/legal assistance	AgentBot in HalluDetect	Role-routed legal intake, advice, and drafting over top-4 retrieved legal chunks
Mobile automation	EcoAgent, MobileAgent, AppAgent-Pro	Action execution, verification, SOP-guided control, proactive multi-app mining
Building operations	OptAgent	Configuration, simulation, control, analysis, comparison across building/HVAC/DER/grid
Physical robotics	AgentRob	Forum-mediated command extraction and robot primitive dispatch

In chemistry, Atreya focuses on bench-to-literature tasks that are difficult to express as one-shot search queries. It retrieves approved drugs by disease, target–drug relations, molecular records by SMILES or InChI Key, tissue ontology entries, USAN stem approvals, and the top 50 approved drugs as CSV, while rendering molecular structure images inline (Sharma et al., 2021).

In insurance, Axlerod is deliberately not a consumer chatbot. It is an agent-assist system over approximately 730,000 personal and commercial policies and a ~400 MB documentation corpus, designed for policy number identification, AutoPay enrollment checks, bill plan identification, covered vehicle lookup, and document retrieval (Bradley et al., 24 Dec 2025). In legal assistance, HalluDetect’s AgentBot uses shared retrieval plus role specialization so that the paralegal gathers facts, the lawyer gives remedies and next steps, and the drafter produces notices or complaints with placeholders such as [Address] when required facts are absent (Anaokar et al., 15 Sep 2025).

In mobile and GUI automation, the emphasis shifts from question answering to grounded action. MobileAgent supports PRESS_BACK, PRESS_HOME, PRESS_ENTER, TYPE, DUAL_POINT, SCROLL, STATUS_TASK_COMPLETE, and STATUS_TASK_IMPOSSIBLE, and adds Structured Instructions for privacy-sensitive confirmation flows (Ding, 2024). EcoAgent extends this actional view with edge-cloud collaboration, while AppAgent-Pro emphasizes proactive information integration across apps rather than only task execution (Yi et al., 8 May 2025, Zhao et al., 26 Aug 2025).

OptAgent applies the paradigm to intelligent buildings. It couples an agentic layer of concierge, orchestrator, and 11 specialist agents with a physics-informed environment spanning building thermal dynamics, HVAC, DER assets, and grid signals, exposing 72 MCP tools for configuration, simulation, analysis, and comparison (Jiang et al., 27 Jan 2026). AgentRob extends AgentBot into embodied robotics, where natural-language commands from forum posts are decomposed into robot primitives such as act_move, act_hello, act_heart, and act_backflip for Unitree Go2 or G1 platforms (Liu et al., 14 Feb 2026).

5. Evaluation regimes and empirical performance

Evaluation practices vary sharply across the literature. Some systems are proofs of concept without formal benchmarking. Atreya reports no formal quantitative evaluation or user study; the workflow diagram contains the label “Top 5 result (Accuracy 100%),” but the paper provides no method, sample size, or metric definition (Sharma et al., 2021). AppAgent-Pro is presented as a system demonstration with three scenarios and no quantitative baselines (Zhao et al., 26 Aug 2025). BMW Agents similarly offers demonstrations rather than benchmarked metrics (Crawford et al., 2024).

Where evaluations are reported, they are heterogeneous and domain-specific. Axlerod reports 93.18% overall accuracy with 1,118 correct [out](https://www.emergentmind.com/topics/outer-automorphism-out) of 1,200 queries across four task types, and a reduction in average task time from 7.55 seconds to 5.13 seconds, i.e., 2.42 seconds, in a small comparison with two agents (Bradley et al., 24 Dec 2025). HalluDetect reports that AgentBot achieves the lowest hallucinations per turn and the highest token accuracy among the tested chatbot architectures: HPT-1 0.4159, HPT-2 0.4279, TokAcc-1 96.13%, TokAcc-2 96.38%, with 46.67% of chats containing at least one hallucination (Anaokar et al., 15 Sep 2025).

Mobile and orchestration systems are evaluated with task-completion or success metrics. On AitW, MobileAgent’s Llama 2+SOP reaches an overall task completion score of 66.92, compared with 65.43 for Llama 2, while preserving the same 0.071 s/inference and response length 21 as the non-SOP baseline (Ding, 2024). On AndroidWorld, EcoAgent with OS-Atlas-Pro 4B + Qwen2-VL-2B achieves SR 27.57, ST 5.33, MC 1.53, and MT 3,240, compared with M3A at SR 28.44, MC 13.39, and MT 87,469; the paper summarizes this as approximately 97% of M3A’s success rate with about 1/27 of the token usage (Yi et al., 8 May 2025).

At the orchestration level, AOrchestra reports strong pass-rate improvements across GAIA, Terminal-Bench, and SWE-Bench-Verified. With Gemini-3-Flash, it reaches 80.00 pass@1 on GAIA, 52.86 on Terminal-Bench, and 82.00 on SWE-Bench-Verified, with an average pass@1 of 71.62; the paper states a 16.28% relative improvement against the strongest baseline under the same model pairing (Ruan et al., 3 Feb 2026). AgentOpt, while not itself an AgentBot, contributes an optimization framework for such systems and reports that at matched accuracy the cost gap between the best and worst model combinations can reach 13–32×, while Arm Elimination reduces evaluation budget by 24–67% relative to brute-force search on three of four tasks (Hua et al., 7 Apr 2026).

Benchmark papers expose current limitations. AutomationBench evaluates cross-application API-based agents on 600 public tasks and 600+ private evaluation tasks across 47 apps and about 500 API endpoints, with score

$S = \frac{\sum_{i=1}^{T} S_i}{T}.$

Even its best listed model reaches only 9.9% pass on the private set, which the benchmark presents as evidence that cross-application coordination, autonomous API discovery, and policy adherence remain difficult (Shepard et al., 21 Apr 2026). By contrast, AgentA/B shows that agent populations can already produce statistically analyzable behavioral signals in web experimentation; in its Amazon study, the treatment produced more purchases than control among agents, with $\chi^2(1)=5.51$, $T_{t+1} = \text{PreUnderstand}(S_{t+1}),$0, while several other comparisons were not significant (Wang et al., 13 Apr 2025).

6. Reliability, safety, observability, and open problems

Reliability is handled through a mixture of guardrails, human oversight, verification modules, and post-hoc auditing. MobileAgent inserts Structured Instructions whenever privacy authorization, notification confirmation, or choice selection is required, replacing autonomous execution with structured user confirmation (Ding, 2024). HalluDetect’s AgentBot uses role-scoped prompts and placeholder insertion to avoid fabricated legal specifics, while the separate HalluDetect detector adds expanded retrieval, memory summarization, inconsistency analysis, and a severity filter before counting hallucinations (Anaokar et al., 15 Sep 2025). BMW Agents institutionalizes verification through an independent Verifier that can trigger replanning and re-execution when outputs do not satisfy the original instruction (Crawford et al., 2024).

Safety concerns grow with embodiment and autonomy. AgentRob explicitly frames the risk of “hijacked physical robots” when forum posts become executable robot commands, and recommends permission management, dangerous-command detection, rate limiting, metadata tagging, physical safety perimeters, and human-in-the-loop checkpoints for high-risk actions (Liu et al., 14 Feb 2026). OptAgent addresses a different safety regime: comfort bounds, actuator limits, SOC ranges, ramp-rate caps, shadow-mode execution, and fallbacks to baseline HVAC scheduling or DER constraints in live building operation (Jiang et al., 27 Jan 2026).

Observability has become a parallel research theme because AgentBots themselves can be detected. MARK achieves 97% accuracy, precision 0.972, recall 0.971, and F1 0.971 by fusing TLS, HTTP, and browser-interaction fingerprints to distinguish six agent frameworks, humans, and crawlers (Kang et al., 18 Jun 2026). A related honeysite study reports 0.993 accuracy and 0.993 F1 when IP, TLS, HTTP, and browser features are combined, and argues that stealth and anti-detection mechanisms often increase detectability rather than decrease it (Fayolle et al., 29 Jun 2026). BOTracle likewise finds that a behavior-only DGCNN over website traversal graphs achieves Accuracy 0.9845, Precision 0.9791, Recall 0.9833, and F1 0.9812, supporting the view that behavioral traces are hard to spoof sustainably at scale (Kadel et al., 2024).

Common limitations recur across the literature. Rule-based conversational systems are brittle under synonymy, ambiguity, and multi-turn context (Sharma et al., 2021). Tool-oriented LLM systems often omit detailed retrieval-ranking, NER, or governance mechanisms even when operating in regulated domains (Bradley et al., 24 Dec 2025). GUI agents remain vulnerable to long-horizon branching tasks, dynamic UI states, and domain shift beyond their training distribution (Yi et al., 8 May 2025, Zhao et al., 26 Aug 2025). Benchmarks indicate that cross-application policy-following automation remains far from solved (Shepard et al., 21 Apr 2026). A plausible implication is that future AgentBot work will continue to converge on a small set of themes already visible here: narrower tool contracts, stronger verifier loops, explicit context management, model-routing under cost constraints, and richer human-in-the-loop policies rather than attempts at unconstrained full autonomy.