EnterpriseLab: A Full-Stack Platform for developing and deploying agents in Enterprises

Ankush Agarwal; Harsh Vishwakarma; Suraj Nagaje; Chaitanya Devaguptapu

cs.AIMar 23, 2026

EnterpriseLab: A Full-Stack Platform for developing and deploying agents in Enterprises

Authors: Ankush Agarwal, Harsh Vishwakarma, Suraj Nagaje, Chaitanya Devaguptapu

Organizations: Fujitsu Research India.

Abstract

Deploying AI agents in enterprise environments requires balancing capability with data sovereignty and cost constraints. While small language models offer privacy-preserving alternatives to frontier models, their specialization is hindered by fragmented development pipelines that separate tool integration, data generation, and training. We introduce EnterpriseLab, a full-stack platform that unifies these stages into a closed-loop framework. EnterpriseLab provides (1) a modular environment exposing enterprise applications via Model Context Protocol, enabling seamless integration of proprietary and open-source tools; (2) automated trajectory synthesis that programmatically generates training data from environment schemas; and (3) integrated training pipelines with continuous evaluation. We validate the platform through EnterpriseArena, an instantiation with 15 applications and 140+ tools across IT, HR, sales, and engineering domains. Our results demonstrate that 8B-parameter models trained within EnterpriseLab match GPT-4o's performance on complex enterprise workflows while reducing inference costs by 8-10x, and remain robust across diverse enterprise benchmarks, including EnterpriseBench (+10%) and CRMArena (+10%). EnterpriseLab provides enterprises a practical path to deploying capable, privacy-preserving agents without compromising operational capability.

Figures & tables

Figure 1: Full Stack Platform for Developing Enterprise Agents: The EnterpriseLab modules (1, 2, and 3) collaborate to create specialized agents, which are then deployed to the agentic environment for end-user interaction.

Table 2

Model	LLM	Tool
	Eval	Eval
Qwen3-8B AGRPO (original env)	0.50	0.20
Qwen3-8B AGRPO (modified env)	0.43	0.15
+ 200 samples incremental training	0.48	0.18

Table 3: Adaptation to environment changes on EnterpriseBench. Model performance before and after API schema modifications, with and without incremental training.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Score
GPT-4o	0.47
ToolAce	0.41
XLAM-2-70B	0.40
Qwen3-4B (Base)	0.27
Qwen3-4B (SFT)	0.32
Qwen3-4B (Agentic GRPO)	0.38

Appendix

Table 4: EnterpriseBench results evaluated by GPT-4o. Scores represent execution scores on enterprise task completion.

Model	Score
GPT-4o	0.44
ToolAce	0.39
XLAM-2-70B	0.39
Qwen3-4B (Base)	0.25
Qwen3-4B (SFT)	0.31
Qwen3-4B (Agentic GRPO)	0.36

Appendix

Table 5: EnterpriseBench results evaluated by Claude-4.5-Sonnet. Scores represent execution scores on enterprise task completion.

Model	Input	Output
	($ per 1M tokens)	($ per 1M tokens)
GPT-4o	5.00	15.00
Claude-3.5-Sonnet	3.00	15.00
Gemini-2.5 Pro	1.25	10.00
Qwen3-8B Agentic GRPO	0.50–1.00

Appendix

Table 6: Inference costs via AWS Bedrock for proprietary models and self-hosted deployment for our model on AWS g5.2xlarge instances.

Benchmark	Domain Focus	Multi-App Flow	Dynamic Data	Enterprise Constraints	Key Limitation
AgentBench ( Liu et al., 2024 )	General Reasoning	✗	✗	✗	No Enterprise Context
WebArena ( Zhou et al., )	General Web UI	✗	✓ (Live Sites)	✗	Public Web, No Authorization
SWE-bench ( Yang et al., )	Software Eng.	✗ (GitHub only)	✗ (Repo Snapshot)	✗	Coding Domain Only
WorkArena ( Drouin et al., 2024 )	IT/ServiceNow	✗ (Single Platform)	✓ (Live UI)	✓	Single SaaS Ecosystem
CRMArena ( Huang et al., 2025a )	CRM	✗ (Salesforce)	✗ (Static)	✓	Single Domain (CRM)
Tau-bench ( Yao et al., 2025 )	Customer Service	✗	✓ (Stateful)	✗	Limited Tool Diversity

Appendix

Table 7: Comparison of EnterpriseArena with existing agent benchmarks. EnterpriseArena uniquely targets multi-application enterprise orchestration with dynamic data, distinguishing it from single-domain (CRM, Code) or static benchmarks.

Domain	Task Description	Task Category	Tools Used
Software Engineering	Create a project named ’Nexus’, add frontend, backend, database modules, and create issues in each module.	Multi-step CRUD	create_project , create_module (×3), create_issue (×3), send_channel_message
HR & Recruitment	Read job description from Documents, fetch shortlisted resumes, identify top candidates, and coordinate interview scheduling.	Search & Orchestration	list_files , read_file_content , search_files , get_employees , send_email (×2)
IT & Project Mgmt	Announce MCP agent online in general channel, create MCP-adapters issue in Sprint 2, verify it appears in module.	Communication & CRUD	send_channel_message , create_issue , list_issues , create_cycle
CRM & Business	Create customer TechCorp, add three products to catalog, generate sales order, create invoice, email confirmation.	Multi-entity Workflow	create_customer , create_product (×3), create_order , create_invoice , send_email
DevOps	Search for repository, fork to personal namespace, create feature branch, push config files, open merge request.	Version Control	search_repositories , fork_repository , create_branch , push_files , create_merge_request
File & Storage	List project documents, create Finance folder in OwnCloud, upload monthly reports, share folder info via channel.	Data Organization	list_files , create_folder , upload_file (×2), get_storage_info , send_channel_message

Appendix

Table 8: Examples of complex multi-step tasks in EnterpriseArena demonstrating cross-functional orchestration with specific tool sequences.

Domain	%	Category	%
IT & Project Mgmt	22	CRUD Operations	35
Software Engineering	18	Search & Orchestration	28
Human Resources	15	Multi-entity Workflow	18
CRM & Business	14	Version Control	12
DevOps	12	Cross-functional	7
Communication	10

Appendix

Table 9: Task distribution across enterprise domains and workflow categories.

Benchmark	Train (Size)	Test (Size)
EnterpriseArena	500	500
EnterpriseBench	700	500
CRMArena	600	500 (Function Calling)
Tau-Bench	300 (Retail), 300 (Airline)	115 (Retail), 50 (Airline)

Appendix

Table 10: Train/Test size for Experimental setting

Sym.	Meaning	Sym.	Meaning
$\pi_{\theta}$	Policy model (LLM)	$G$	Group size (rollouts/query)
$\theta$	Trainable params (LoRA)	$g$	Trajectory index $1\dots G$
$\pi_{\text{ref}}$	Frozen reference policy	$\mathcal{T}_{q}$	Trajectory set for query $q$
$\mathcal{D}$	Training dataset	$h$	History (msgs buffer)
$Q$	Batch sampled from $\mathcal{D}$	$T_{\max}$	Max tool steps limit
$q$	Single query/task	$z$	Thought tokens ( <think> )

Appendix

Table 11: Notation for Agentic GRPO

Profession	Gender	Age
Software Engineer	M	25
Senior Engineer	F	29
Business Development Representative	F	26
Sales Manager	F	35
IT Engineer	M	29
IT Security Head	M	32

Appendix

Table 12: Domain Expert Demographics: Professional roles, gender, and age distribution of the nine domain experts who validated EnterpriseArena tasks and environment specifications.

Panel of Figure 2 — Figure 2: Domain Expert Validation in EnterpriseArena. Domain experts from all benchmark domains evaluate the fidelity of the created environment and tasks. This example shows screenshots of MS form for different steps a domain expert completes during the validation process.

Category	MCP Server	# Tools	Representative Tools
Communication	RocketChat	12	send_user_message , send_channel_message , get_channels
Communication	Mail System	8	send_email , read_email , search_emails , draft_email
Development	GitLab MCP	22	create_merge_request , create_issue , push_files , execute_graphql
Operations & IT	Zammad	10	create_ticket , assign_agent , update_status , add_note
Operations & IT	Plane (Jira)	14	create_issue , add_milestone , assign_task , update_state
Human Resources	Frappe HR	14	get_employee , create_leave_request , get_attendance , get_payroll

Appendix

Table 13: MCP Server ecosystem in EnterpriseArena with representative tools per server.

Tool Name	Description & Key Parameters
send_user_message	Send direct message to user (e.g., ’john.doe’, ’@john.doe’). Params: channel (string, required), message (string, required)
send_channel_message	Send message to channel (e.g., ’general’, ’#general’). Params: channel (string, required), message (string, required)
get_channels	List all available channels. Params: None
get_channel_messages	Get recent messages from a channel. Params: channel (string, required), count (integer, optional)
create_channel	Create new public or private channel. Params: name (string, required), description (string, optional), private (boolean, optional)
get_user_info	Get information about current authenticated user. Params: None

Appendix

Table 14: Complete RocketChat MCP Server Tool Catalog

Tool Name	Description & Key Parameters
send_email	Send email with attachments (supports plain text, HTML, and multipart formats). Params: to (string, required), subject (string, required), body (string, required), attachments (array, optional)
draft_email	Create draft email without sending. Params: to (string, required), subject (string, required), body (string, required)
read_email	Retrieve complete email content by ID with attachment info. Params: email_id (string, required)
search_emails	Search emails by subject, sender, or date range. Params: query (string, required), from_date (string, optional), to_date (string, optional)
list_folders	List all mail folders (inbox, sent, drafts, etc.). Params: None
move_email	Move email to specified folder. Params: email_id (string, required), folder (string, required)

Appendix

Table 15: Complete Mail System MCP Server Tool Catalog

Tool Name	Description & Key Parameters
Repository Management
create_repository	Create new GitLab project. Params: name (string, required), description (string, optional), visibility (string, optional), initialize_with_readme (boolean, optional)
search_repositories	Search for GitLab projects. Params: search (string, required), page (number, optional), per_page (number, optional, max 100)
fork_repository	Fork project to your account or namespace. Params: project_id (string, optional), namespace (string, optional)
get_repository_tree	Get repository file tree structure. Params: project_id (string, optional), path (string, optional), ref (string, optional)
File Operations

Appendix

Table 16: GitLab MCP Server Tool Catalog (Selected Core Tools)

Tool Name	Description & Key Parameters
Zammad MCP (IT Ticketing)
create_ticket	Create new support ticket. Params: title (string, required), description (string, required), priority (string, optional), category (string, optional)
get_ticket	Retrieve ticket details by ID. Params: ticket_id (string, required)
update_ticket	Update ticket status, priority, or assignee. Params: ticket_id (string, required), status (string, optional), priority (string, optional), assignee_id (string, optional)
search_tickets	Search tickets by criteria. Params: query (string, required), status (string, optional), assignee (string, optional)
add_note	Add internal or public note to ticket. Params: ticket_id (string, required), note (string, required), internal (boolean, optional)

Appendix

Table 17: Zammad & Plane MCP Server Tool Catalogs

Tool Name	Description & Key Parameters
Frappe HR MCP
get_employees	Retrieve employee list with filters. Params: department (string, optional), designation (string, optional), status (string, optional)
get_employee	Get employee details by ID. Params: employee_id (string, required)
create_employee	Onboard new employee. Params: first_name (string, required), last_name (string, required), email (string, required), department (string, required)
get_departments	List organization departments. Params: None
get_attendance	Retrieve attendance records. Params: employee_id (string, optional), from_date (string, optional), to_date (string, optional)

Appendix

Table 18: Frappe HR, OwnCloud, Mongoose, & Dolibarr MCP Server Catalogs

Tool Name	Description & Key Parameters
File System MCP
read_file	Read file contents. Params: path (string, required)
write_file	Write data to file. Params: path (string, required), content (string, required)
list_directory	List directory contents. Params: path (string, required)
create_directory	Create new directory. Params: path (string, required)
Web Browser MCP

Appendix

Table 19: Utility MCP Server Tool Catalog (File System, Browser, Bash)

Enterprise AI Agents

See all 15 papers

SplitAgent: A Privacy-Preserving Distributed Architecture for Enterprise-Cloud Agent Collaboration
Mar 9, 2026Jianshu SheAI Agent Security Multi-Agent Collaboration
MBZUAI
EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings
Mar 13, 2026Shiva Krishna Reddy Malay, Shravan Nayak, Jishnu Sethumadhavan Nair +6AI Agent Benchmarks LLM Agent Planning
ServiceNow Research · Mila - Quebec AI Institute · Université de Montréal
Beyond Autonomy: A Dynamic Tiered AgentRunner Framework for Governable and Resilient Enterprise AI Execution
May 11, 2026Kai Pan, Rong HouLLM Agent Orchestration AI Agent Safety
A2A Lab

Mar 9, 2026J/K move · Enter open · S save

SplitAgent: A Privacy-Preserving Distributed Architecture for Enterprise-Cloud Agent Collaboration

Jianshu She

MBZUAI

AI Agent Security Multi-Agent Collaboration

Open paper AskPDF arXiv

Enterprise adoption of cloud-based AI agents faces a fundamental privacy dilemma: leveraging powerful cloud models requires sharing sensitive data, while local processing limits capability. Current agent frameworks like MCP and A2A assume complete data sharing, making them unsuitable for enterprise environments with confidential information. We present SplitAgent, a novel distributed architecture that enables privacy-preserving collaboration between enterprise-side privacy agents and cloud-side reasoning agents. Our key innovation is context-aware dynamic sanitization that adapts privacy protection based on task semantics -- contract review requires different sanitization than code review or financial analysis. SplitAgent extends existing agent protocols with differential privacy guarantees, zero-knowledge tool verification, and privacy budget management. Through comprehensive experiments on enterprise scenarios, we demonstrate that SplitAgent achieves 83.8% task accuracy while maintaining 90.1% privacy protection, significantly outperforming static approaches (73.2% accuracy, 79.7% privacy). Context-aware sanitization improves task utility by 24.1% over static methods while reducing privacy leakage by 67%. Our architecture provides a practical path for enterprise AI adoption without compromising sensitive data.

Abstract

Enterprise AI Agents

Explore similar work

SplitAgent: A Privacy-Preserving Distributed Architecture for Enterprise-Cloud Agent Collaboration

EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings

Beyond Autonomy: A Dynamic Tiered AgentRunner Framework for Governable and Resilient Enterprise AI Execution