Agentic AI reference agents: 120 systems under test

There are 132 reference agents. A1-A12 are small single-pattern agents: no-LLM rules, a single LLM call, RAG, prompt chains, handoffs, ReAct with human approval, MCP tools, low-code n8n, multi-agent crews, coding teams, evaluator-optimizer guardrails and durable approval workflows. A13-A25 are larger, production-style systems: a LangGraph supervisor, LangGraph map-reduce, plan-and-execute with replanning, a multi-server MCP agent, multi-hop RAG with verified citations, a CrewAI hierarchical crew, a long-running assistant with memory across processes, an AutoGen SelectorGroupChat debate, an AutoGen negotiation, a guarded browser agent on a local mock site, an n8n orchestrator with sub-workflows, a CrewAI Flows KYC pipeline and an AutoGen Swarm incident triage with human handoff. A26-A70 are practical enterprise agents (banking, insurance, healthcare admin, HR, SRE, analytics and more), each read-only or with approval-gated writes to synthetic mocks, with real unit tests and three recorded runs. A71-A81 are coverage agents (phase 5) in plain Python with OpenAI-compatible function calling and native OpenTelemetry GenAI traces; A82-A98 (phase 6) add fixed and seeded-defect variants so that every agentest check has both a passing and a failing example (agentest v3.2.1: seed-aware catch-rate 90/90). Fixed agents keep writes read-only or approval-gated; seeded-defect agents (A90-A98) intentionally perform unrestricted writes so agentest has real failing cases. A99-A101 (phase 7) are genuine multi-agent desks (motor claims, Hinglish banking disputes, flight disruption) recorded against agentest v3.2.1 then re-audited on v3.3.0; A102-A104 (phase 7b) add hotel overbooking, telecom outage and library ILL desks built only after v3.3.0 rules were frozen (see agentest docs/BLIND-TEST.md). A105-A110 (phase 8) are genuine decision desks where bias and explanations matter (home-loan pre-approval, resume screening, motor claims, ED triage, tenant screening, Hindi/English/Tamil content moderation), each recorded with counterfactual variants that change one protected attribute. A111-A113 (phase 9) add a merit-scholarship award desk, a travel-expense reimbursement desk and a pharmacy refill desk, built only after the agentest v3.4 rules were frozen and blind-tested against agentest v3.4.0 (manual audit first; results in agentest docs/BLIND-TEST.md, round 3). A114-A116 (phase 10) add a vendor invoice payment desk, a SIM-swap fraud desk (clean control) and a privileged-access request desk, built only after the round-4 rules freeze and blind-tested against agentest v3.4 (round 4). A117-A120 (phase 11) add a letter-of-credit document desk, a full-and-final settlement desk (clean control), a multilingual card-dispute desk and a vendor onboarding desk, built only after the round-5 rules freeze and blind-tested against agentest v4.0 (round 5); A119 runs on Gemma 3 4B and A120 on Moonlight-16B-A3B, a small open Kimi-family model (not Kimi K2). A121-A126 (phase 12) add a medical-records release desk (privacy), a railway lost-property desk (clean control), a multilingual crop-insurance desk (English/Hindi/Telugu), a SaaS account-takeover desk (security, injected instruction), a self-help-group microloan desk and an English/Marathi electricity bill desk, built only after the round-6 rules freeze and blind-tested against agentest v4.0 (round 6); A121-A122 run on Qwen2.5-7B, A123-A124 on Gemma 3 4B and A125-A126 on the small open Kimi-family (Moonlight-16B-A3B Q2_K), not K2. A127-A132 (phase 13) add a mobile number-portability desk (clean control), a school RTE-quota admission desk (privacy), a PDS ration-card grievance desk (English/Tamil/Bengali), a UPI refund-dispute desk (security, injected instruction), an English/Gujarati LPG subsidy desk and an English/Punjabi e-challan dispute desk, built only after the round-7 rules freeze and blind-tested against agentest v4.0 (round 7) on the same three models. They are the systems under test for the open Agentic AI Test Checklist 110 catalogue, and each one lists the catalogue tests that apply to it.

This repo contains code, not model weights. Every LLM agent uses Qwen2.5-7B-Instruct (GGUF Q4_K_M, about 4.7 GB) running locally on CPU through llama.cpp behind a small OpenAI-compatible server (common/gguf_llm_server.py). The first runs used Qwen2.5-1.5B-Instruct (common/local_llm_server.py); those runs are kept in aNN_*/proof/baseline-1.5B/ as a comparison, because their failures are useful test material, so everything runs at zero cost with no API keys. The back-ends are fictional mocks with .example addresses and canary strings.

Browse every recorded run: viewer Space. It is a static page, because hosted Gradio/Docker Spaces now need a paid plan. To run the agents live, use your own machine (python app.py starts a Gradio UI with one tab per agent).

The 120 agents

ID	Folder	Use case	Framework	LLM	Phase	Status (7B)	Baseline 1.5B	Tests that apply
A1	a01_rule_router	Rule-based ticket router (no LLM)	n8n (Code + Switch nodes), plus a pure-Python twin	no	1	partial 3/4 (proof)	-	10
A2	a02_llm_classifier	Single-call LLM classifier/extractor	Plain OpenAI-compatible API (one call)	yes	1	working 3/3 (proof)	partial 1/3	17
A3	a03_rag_faq	RAG FAQ assistant	LlamaIndex (VectorStoreIndex + bge-small embeddings)	yes	1	working 3/3 (proof)	partial 2/3	46
A4	a04_prompt_chain	Prompt-chain content pipeline (draft, check, translate, review)	LangGraph	yes	3	working 3/3 (proof)	working 3/3	52
A5	a05_support_handoffs	Customer-support router with refunds	OpenAI Agents SDK (openai-agents, handoffs, function tools)	yes	2	working 3/3 (proof)	working 3/3	96
A6	a06_react_assistant	ReAct calendar/email assistant	LangGraph (create_react_agent + interrupt)	yes	1	working 3/3 (proof)	working 3/3	96
A7	a07_mcp_data_agent	MCP-connected enterprise data agent (read-only)	AutoGen AgentChat + McpWorkbench, local MCP server over stdio	yes	2	partial 2/3 (proof)	partial 2/3	62
A8	a08_n8n_mcp_agent	Low-code AI agent with MCP tools	n8n AI Agent node + MCP Client Tool + Window memory (self-hosted n8n)	yes	2	partial 2/3 (proof)	partial 2/3	76
A9	a09_research_crew	Research and summarisation crew	CrewAI (Flow + sequential Crew of 3 agents)	yes	3	working 3/3 (proof)	working 3/3	86
A10	a10_coding_team	Orchestrator-workers coding agent	AutoGen AgentChat (RoundRobinGroupChat: planner, coder, CodeExecutorAgent, reviewer)	yes	3	partial 0/3 (proof)	partial 0/3	105
A11	a11_evaluator_optimizer	Evaluator-optimizer with guardrail agent	OpenAI Agents SDK (input_guardrail + generator/critic loop)	yes	2	working 3/3 (proof)	partial 2/3	40
A12	a12_durable_approval	Long-running approval workflow (finance/ops)	LangGraph + SqliteSaver checkpointer + MCP (stdio) + function calls	yes	3	working 3/3 (proof)	working 3/3	99
A13	a13_supervisor_ops	Hierarchical supervisor for customer operations	LangGraph (supervisor node + 3 tool-calling worker agents)	yes	3	working 3/3 (proof)	failed 0/0	97
A14	a14_mapreduce_contracts	Map-reduce contract review	LangGraph Send API (parallel clause reviewers) + reduce	yes	3	partial 2/3 (proof)	failed 0/0	27
A15	a15_plan_execute	Plan-and-execute travel agent with replanning	LangGraph (planner -> executor -> replanner loop)	yes	3	partial 2/3 (proof)	failed 0/0	97
A16	a16_mcp_multiserver	Multi-server MCP supply-chain agent	AutoGen AgentChat AssistantAgent + 2 McpWorkbench (two stdio MCP servers)	yes	3	working 3/3 (proof)	failed 0/0	85
A17	a17_multihop_rag	Multi-hop RAG with verified citations	LangGraph (hop planner -> BM25 retrieve loop -> answer -> citation verifier)	yes	3	partial 1/3 (proof)	failed 0/0	61
A18	a18_crew_hierarchical	Vendor due-diligence crew (hierarchical)	CrewAI Process.hierarchical (manager_llm + financial, compliance, writer agents)	yes	3	working 3/3 (proof)	failed 0/0	97
A19	a19_memory_assistant	Long-running assistant with memory across sessions	LangGraph + SqliteStore (long-term memory), one OS process per session	yes	3	partial 2/3 (proof)	failed 0/0	87
A20	a20_debate_consensus	Credit-committee debate with consensus	AutoGen AgentChat SelectorGroupChat (risk, growth, compliance, moderator)	yes	3	partial 2/3 (proof)	failed 0/0	96
A21	a21_negotiation	Procurement negotiation (buyer vs seller agents)	AutoGen AgentChat RoundRobinGroupChat (buyer + seller, tool-only actions)	yes	3	working 3/3 (proof)	failed 0/0	96
A22	a22_guarded_browser	Guarded browser agent on a local mock shop	OpenAI Agents SDK (function tools) + text browser over a local http.server site	yes	3	partial 2/3 (proof)	failed 0/0	96
A23	a23_n8n_orchestrator	n8n orchestrator calling sub-workflows	n8n (HTTP Request -> local LLM, IF routing, Execute Workflow -> 3 sub-workflows)	yes	3	working 3/3 (proof)	failed 0/0	82
A24	a24_kyc_flow	KYC document pipeline	CrewAI Flows (@start -> @listen -> @router -> branches) + deterministic KYC rules	yes	3	working 3/3 (proof)	failed 0/0	53
A25	a25_incident_swarm	Incident-response triage swarm with human approval	AutoGen AgentChat Swarm (triage -> investigator -> remediator -> handoff to user)	yes	3	working 3/3 (proof)	failed 0/0	96

Enterprise agents A26-A70 (45 added so far, in batches of 5)

Practical multi-step agents for enterprise use cases (banking, insurance, healthcare admin, HR, SRE, and more), built on AutoGen AgentChat RoundRobinGroupChat, AutoGen AgentChat SelectorGroupChat, AutoGen AgentChat Swarm, CrewAI Flows, CrewAI sequential crew, Haystack Pipeline, Hugging Face smolagents ToolCallingAgent, Hugging Face smolagents ToolCallingAgent + OTP and session checks in tools, Hugging Face smolagents ToolCallingAgent + coverage rules in code, Hugging Face smolagents ToolCallingAgent + offer verification in code, Hugging Face smolagents ToolCallingAgent + sender and call-back checks in code, LangGraph, LangGraph StateGraph, LangGraph StateGraph with a conditional edge, LangGraph reflection loop, LlamaIndex Workflow, LlamaIndex Workflows, MCP Python SDK, Microsoft Semantic Kernel ChatCompletionAgent + ExpensePlugin, Microsoft Semantic Kernel ChatCompletionAgent + GSTPlugin, Microsoft Semantic Kernel ChatCompletionAgent + ReturnsPlugin, Microsoft Semantic Kernel ChatCompletionAgent + ScholarshipPlugin, OpenAI Agents SDK, Plain Python durable workflow. Each one has deterministic guards in code, approval-gated writes to synthetic mock back-ends, real offline unit tests, three recorded cases with a JSONL trace, and its own README with a mermaid diagram, the catalogue tests that apply and references. Status is honest: working = 3/3 sanity checks, partial = it runs but at least one case failed (see DEFECTS.md).

ID	Folder	Use case	Framework	Status (7B)	Tests that apply
A26	a26_loan_underwriting	Retail loan underwriting with human sanction	LangGraph (extract -> rules -> explain) + human sanction gate	working 3/3 (proof)	40
A27	a27_health_claims_flow	Health-insurance claim adjudication	CrewAI Flows (@start -> @listen -> @router -> branches) + deterministic policy rules	working 3/3 (proof)	53
A28	a28_prior_auth_agent	Healthcare prior-authorisation agent	OpenAI Agents SDK (function tools) + payer criteria re-check in code	partial 2/3 (proof)	82
A29	a29_hr_onboarding	HR onboarding coordinator with least-privilege provisioning	AutoGen AgentChat RoundRobinGroupChat (planner + provisioner with tools)	working 3/3 (proof)	83
A30	a30_mcp_sre_runbook	SRE on-call agent over MCP	MCP Python SDK (stdio ClientSession) + tool-calling loop on the local LLM	working 3/3 (proof)	85
A31	a31_text_to_sql	Text-to-SQL analytics agent (read-only, self-correcting)	LlamaIndex Workflows (draft -> run -> self-correct -> answer) + SQLite authorizer	working 3/3 (proof)	44
A32	a32_ci_triage_smol	CI failure triage agent	Hugging Face smolagents ToolCallingAgent (5 tools, max 7 steps)	working 3/3 (proof)	83
A33	a33_contract_redline_haystack	Contract redline reviewer against a legal playbook	Haystack Pipeline (InMemoryBM25Retriever -> ChatPromptBuilder -> OpenAIChatGenerator) + playbook rules	working 3/3 (proof)	39
A34	a34_returns_sk	E-commerce returns and refunds agent	Microsoft Semantic Kernel ChatCompletionAgent + ReturnsPlugin (auto function calling)	partial 2/3 (proof)	82
A35	a35_shipment_exception	Shipment exception handler (durable, resumable)	Plain Python durable workflow (JSON checkpoint per step, resumable, idempotent writes)	working 3/3 (proof)	54
A36	a36_reorder_reflection	Inventory reorder planner with a critic loop	LangGraph reflection loop (propose -> deterministic critic -> revise, max rounds)	working 3/3 (proof)	40
A37	a37_support_swarm	Broadband customer-support swarm with handoffs	AutoGen AgentChat Swarm (triage -> billing	tech -> handoff to user)	working 3/3 (proof)
A38	a38_crm_lead_crew	Inbound lead qualification crew	CrewAI sequential crew (researcher with CRM tool -> qualifier -> writer) + scoring and consent rules	working 3/3 (proof)	82
A39	a39_ap_three_way_match	Accounts-payable 3-way match	LangGraph (extract -> match -> route) + deterministic PO/GRN/vendor-master checks	working 3/3 (proof)	39
A40	a40_ar_collections	Accounts-receivable collections agent	OpenAI Agents SDK (function tools) + dunning, dispute and plan policies	working 3/3 (proof)	82
A41	a41_helpdesk_reset_smol	IT helpdesk password reset with OTP verification	Hugging Face smolagents ToolCallingAgent + OTP and session checks in tools	working 3/3 (proof)	82
A42	a42_kyc_onboarding	Bank KYC onboarding checker	LangGraph (extract -> rules -> route) + PAN, name-match, age and sanctions rules	working 3/3 (proof)	39
A43	a43_meeting_scheduler	Cross-time-zone meeting scheduler	OpenAI Agents SDK (function tools) + working-hours slot search in code	partial 2/3 (proof)	82
A44	a44_motor_fnol_haystack	Motor insurance first notice of loss	Haystack Pipeline (BM25 retriever -> ChatPromptBuilder -> OpenAIChatGenerator) + coverage rules	working 3/3 (proof)	39
A45	a45_expense_audit_sk	Expense report auditor	Microsoft Semantic Kernel ChatCompletionAgent + ExpensePlugin (auto function calling)	working 3/3 (proof)	82
A46	a46_rfq_award_flow	RFQ evaluation and purchase-order award	CrewAI Flows (@start extract -> @listen score -> @router award	rebid)	working 3/3 (proof)
A47	a47_postmortem_writer	Incident postmortem writer	LlamaIndex Workflow (draft -> check -> revise once) + grounding, roster and blameless checks	working 3/3 (proof)	40
A48	a48_dsar_privacy	Data-subject (DSAR) privacy request handler	LangGraph (classify -> verify identity -> inventory -> act)	working 3/3 (proof)	39
A49	a49_pharmacy_refill	Pharmacy prescription refill assistant	OpenAI Agents SDK (function tools) + refill rules in code	working 3/3 (proof)	82
A50	a50_moderation_selector	Content moderation panel	AutoGen AgentChat SelectorGroupChat (analyst -> policy checker -> enforcer) + policy matrix	working 3/3 (proof)	39
A51	a51_churn_retention_crew	Telecom churn-retention crew	CrewAI sequential crew (usage analyst with tool -> offer planner -> message writer) + offer caps in code	working 3/3 (proof)	82
A52	a52_vendor_bank_change_smol	Vendor bank-detail change verifier	Hugging Face smolagents ToolCallingAgent + sender and call-back checks in code	working 3/3 (proof)	82
A53	a53_trial_prescreen_haystack	Clinical-trial pre-screener	Haystack Pipeline (BM25 over protocol criteria -> ChatPromptBuilder -> OpenAIChatGenerator) + criteria in code	working 3/3 (proof)	39
A54	a54_gst_itc_sk	GST input-tax-credit checker	Microsoft Semantic Kernel ChatCompletionAgent + GSTPlugin (auto function calling)	working 3/3 (proof)	82
A55	a55_field_dispatch	Field-service dispatcher with a safety branch	LangGraph StateGraph with a conditional edge (triage -> safety	match -> assign)	working 3/3 (proof)
A56	a56_travel_policy_booker	Corporate travel booker	OpenAI Agents SDK (function tools) + travel policy in code	working 3/3 (proof)	82
A57	a57_resume_screen_flow	Resume screener with a counterfactual fairness check	CrewAI Flows (@start redact -> @listen extract x2 -> @listen score -> @router)	working 3/3 (proof)	39
A58	a58_change_advisory_rr	Change advisory board	AutoGen AgentChat RoundRobinGroupChat (risk assessor -> reviewer -> scheduler) + change policy in code	partial 2/3 (proof)	39
A59	a59_price_match_smol	Retail price-match agent	Hugging Face smolagents ToolCallingAgent + offer verification in code	working 3/3 (proof)	82
A60	a60_clinic_triage_wf	Clinic symptom triage and booking	LlamaIndex Workflow (red-flag screen -> route -> book	emergency)	working 3/3 (proof)
A61	a61_outage_comms	Power-outage customer communications	LangGraph StateGraph (fetch -> draft -> check -> revise once -> send)	working 3/3 (proof)	39
A62	a62_customs_hs_haystack	Customs HS-code classifier	Haystack Pipeline (BM25 over a tariff extract -> ChatPromptBuilder -> OpenAIChatGenerator) + duty in code	partial 2/3 (proof)	39
A63	a63_scholarship_sk	Scholarship eligibility officer	Microsoft Semantic Kernel ChatCompletionAgent + ScholarshipPlugin	working 3/3 (proof)	82
A64	a64_card_dispute	Card dispute agent	OpenAI Agents SDK (function tools) + dispute rules in code	working 3/3 (proof)	82
A65	a65_research_brief_crew	Market-research brief with citation verification	CrewAI sequential crew (researcher with search tool -> analyst -> editor) + quote verification in code	partial 2/3 (proof)	39
A66	a66_supplier_risk_wf	Supplier risk monitor	LlamaIndex Workflow (fetch news -> classify signals -> score -> act)	working 3/3 (proof)	39
A67	a67_saas_license_graph	SaaS licence reclamation	LangGraph StateGraph (inventory -> propose -> check) + eligibility in code	partial 2/3 (proof)	39
A68	a68_warranty_claim_smol	Appliance warranty claim agent	Hugging Face smolagents ToolCallingAgent + coverage rules in code	working 3/3 (proof)	82
A69	a69_meter_billing_sdk	Electricity billing correction	OpenAI Agents SDK (function tools) + re-billing on a telescopic tariff in code	partial 2/3 (proof)	82
A70	a70_soc_containment_selector	SOC containment panel	AutoGen AgentChat SelectorGroupChat (triage -> intel -> responder) + tier-based approver policy	working 3/3 (proof)	39

Coverage agents A71-A81 (added 2026-10-27)

Each ships an A2A AgentCard, an MCP tools.json and OTLP GenAI traces in proof/, and nothing agent-specific is needed to test it. Status: working means every demo sanity check and every agentest check the agent was built to exercise passes; in progress means at least one still fails (the failure is kept and described in the agent README). The last column is the number of agentest v3.1 checks (of 130) that give a pass or fail on the recorded traces.

ID	Folder	Use case	Framework	Status (7B)	agentest checks decided
A71	a71_order_fulfilment_saga	Order fulfilment saga	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 5/5 (proof)	37
A72	a72_tenant_analytics_export	Multi-tenant analytics export	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 1/4 (proof)	37
A73	a73_web_research_fetcher	Web research fetcher	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 3/4 (proof)	28
A74	a74_file_workspace_agent	File workspace agent	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 3/5 (proof)	42
A75	a75_crm_bulk_updater	CRM bulk updater	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 1/5 (proof)	37
A76	a76_concierge_memory	Office concierge with memory	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	working 14/14 (proof)	38
A77	a77_pr_review_bot	Pull-request review bot	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 4/5 (proof)	33
A78	a78_iam_access_desk	IAM access desk (L1 -> L2)	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 4/4 (proof)	41
A79	a79_invoice_image_intake	Invoice image intake	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 2/3 (proof)	37
A80	a80_crossborder_payout_saga	Cross-border payout saga	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	in progress 2/4 (proof)	36
A81	a81_inventory_adjuster	Inventory adjuster	Plain Python + OpenAI-compatible function calling (native OTel GenAI traces)	working 3/3 (proof)	29

Status: working means every demo sanity check passed. Partial means the agent runs end to end but the model got at least one case wrong; those failures are kept on purpose, because they are what the tests are for.

Taxonomy

Dimension	Values used
Steps	single-step; multi-step
Reasoning	rule-based (no LLM); LLM workflow (fixed code path); autonomous LLM agent; hybrid
Agent count	1; handoff network; role crew of 2 to 4; supervisor + workers; map-reduce fan-out; selector group chat; swarm; adversarial pair (negotiation)
Tooling	none; function calls; MCP (stdio and Streamable HTTP, single and multi-server); code execution; text browser on a local mock site; n8n sub-workflows
Memory/RAG	none; session; RAG; multi-hop RAG with citation verification; persistent checkpointed state; long-term memory store across processes
Human in the loop	none; review; approval before write actions; swarm handoff to a human
Frameworks	n8n (AI Agent, Execute Workflow), LangGraph (StateGraph, Send, SqliteSaver, SqliteStore), CrewAI (sequential, hierarchical, Flows), AutoGen AgentChat (RoundRobin, SelectorGroupChat, Swarm, McpWorkbench), OpenAI Agents SDK, LlamaIndex, MCP Python SDK

Quickstart (one command)

git clone https://huggingface.co/srinivaspadmanabhuni/agentic-ai-reference-agents && cd agentic-ai-reference-agents
bash setup.sh               # Linux/macOS: venv + deps + 7B GGUF (~4.7 GB, free, no login) + smoke check
bash setup.sh --no-model    # deps + 255 offline unit tests only (no download, about 5 min)
bash setup.sh --agent a16   # live smoke check of another agent (default a24, about 2 min on CPU)

Windows: powershell -ExecutionPolicy Bypass -File .\setup.ps1 (same steps; -NoModel, -Agent a16). The smoke check runs the offline unit tests, then verify.py --agent a24, which starts the local model server, reruns the three recorded cases into verify-out/ and compares them with proof/ (tested on a fresh clone, 26 Oct 2026: 67 unit tests passed, a24 3/3 with 0 differences, about 4 min with the model already downloaded). Needs Python 3.11 or 3.12 and about 6 GB of free RAM. The 785 "skipped" entries in the pytest output are the A1-A12 tests/ index files, which list mapped test IDs and are not checks.

Run locally (manual steps)

git clone https://huggingface.co/srinivaspadmanabhuni/agentic-ai-reference-agents && cd agentic-ai-reference-agents
python3.12 -m venv .venv && . .venv/bin/activate && pip install -r requirements.txt
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
hf download bartowski/Qwen2.5-7B-Instruct-GGUF Qwen2.5-7B-Instruct-Q4_K_M.gguf --local-dir gguf   # ~4.7 GB, free, no login
python common/gguf_llm_server.py &       # default model, serves :8008 (needs ~6 GB free RAM)
# fallback for small machines: python common/local_llm_server.py &  with REF_MODEL=Qwen/Qwen2.5-1.5B-Instruct (~3 GB, transformers)
cd a06_react_assistant && python run_demo.py   # any agent folder; each writes proof/run.md + proof/run.jsonl
cd .. && python app.py                    # optional Gradio UI, one tab per agent
pytest -q                                  # A13+ unit tests + common/ tests run for real; A1-A12 stubs report 'skipped'

For the n8n agents (A1, A8 and A23), install Node 22 and run npm i n8n in a folder, set N8N_DIR to it, then run ./a01_rule_router/run_n8n.sh, python a08_n8n_mcp_agent/run_demo.py or python a23_n8n_orchestrator/run_demo.py. n8n runs self-hosted, with no account.

Environment: REF_GGUF points gguf_llm_server.py at another GGUF file. REF_MODEL sets the model name shown in the proofs; with the fallback server it also chooses the transformers model, and with the fallback server also set REF_MODEL=Qwen/Qwen2.5-1.5B-Instruct before running agents so proofs label it correctly and REF_LLM_BASE points the agents at another OpenAI-compatible server.

Genuine multi-agent desks A99-A101 (phase 7, added 2026-10-27)

Fresh multi-agent desks with an orchestrator plus two or more sub-agents (nested invoke_agent spans), several local Qwen2.5-7B calls at zero cost, and MCP-annotated tools. No deliberate defect seeding. Blind-tested against a fresh pip install of agentest tag v3.2.1 in both strict and examples modes; an independent auditor review is in agentest docs/BLIND-TEST.md. Status is honest: working = 3/3 sanity checks; agentest fail counts are real findings (true/false positives called out in the blind-test report).

ID	Folder	Use case	Framework	Status (7B)	agentest v3.2.1 blind
A99	a99_motor_claims_desk	Motor insurance claims triage (intake -> fraud -> payout + ApprovalGate)	Plain Python multi-agent + function calling (native OTel GenAI)	working 3/3 (proof)	37 pass / 5 fail
A100	a100_hinglish_bank_desk	Hinglish retail-banking dispute escalation (triage -> L1 Hindi -> L2 English)	Plain Python multi-agent + function calling (native OTel GenAI)	working 3/3 (proof)	26 pass / 3 fail
A101	a101_flight_disruption_desk	Airline flight-disruption desk (ops -> schedule -> compensation -> rebook)	Plain Python multi-agent + function calling (native OTel GenAI)	working 3/3 (proof)	25 pass / 2 fail

Genuine multi-agent desks A102-A104 (phase 7b, added 2026-10-27)

Built after agentest v3.3.0 rules were frozen; no retune after seeing audit results. Local Qwen2.5-7B; native OTel GenAI.

ID	Folder	Domain	Framework	Status	agentest v3.3.0 (examples)
A102	a102_hotel_overbooking_desk	Hotel overbooking recovery (front-desk -> inventory -> guest-relations)	Plain Python multi-agent + function calling (native OTel GenAI)	working 3/3 (proof)	30 pass / 0 fail
A103	a103_telecom_outage_desk	Telecom outage ticket (NOC -> field -> customer-care)	Plain Python multi-agent + function calling (native OTel GenAI)	working 3/3 (proof)	30 pass / 0 fail
A104	a104_library_ill_desk	Campus library ILL (intake -> partner-search -> fulfillment)	Plain Python multi-agent + function calling (native OTel GenAI)	partial 2/3 (proof)	31 pass / 3 fail (T5-006, T6-003, X-011)

Bias and explainability desks A105-A110 (phase 8, added 2026-10-27)

Genuine multi-agent decision desks for agentest v3.4 (no deliberate defect seeding): an orchestrator plus two sub-agents, 5+ steps per run, several local Qwen2.5-7B calls at zero cost, MCP-annotated tools. Each AgentCard publishes a fairness block; agentest generates the counterfactual inputs from the card's skill examples (one protected attribute changed per variant, marked agentest.input.counterfactual) and every variant was recorded. Status is honest: working = all canonical sanity checks pass; counterfactual runs are recorded as-is with no built-in assertion. agentest column: v3.4 branch (pre-release), examples mode, judge off (semantic facets of T7-007 etc. report not assessed until a judge is configured); fails are real findings, kept on purpose.

ID	Folder	Decision	Framework	Status (7B)	Recorded runs; agentest v3.4-dev
A105	a105_loan_preapproval_desk	Home-loan pre-approval (loan-intake -> underwriter -> offers-clerk)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+18 counterfactual runs; agentest v3.4-dev: 36p/2f T1-008, T7-007
A106	a106_resume_screening_desk	Resume screening (screening-intake -> skills-matcher -> recruiting-coordinator)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4-dev: 34p/0f
A107	a107_insurance_claims_desk	Motor insurance claims (claim-intake -> adjuster -> payments-clerk)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4-dev: 33p/0f
A108	a108_medical_triage_desk	ED triage priority (triage-nurse -> triage-physician -> care-scheduler)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4-dev: 31p/0f
A109	a109_tenant_screening_desk	Tenant screening (leasing-intake -> screening-officer -> leasing-agent)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4-dev: 33p/2f T5-002, X-012
A110	a110_multilingual_moderation_desk	Content moderation in hi/en/ta (mod-intake -> policy-reviewer -> enforcement)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+9 counterfactual runs; agentest v3.4-dev: 31p/4f T1-006, T1-011, T2-003, T7-007

Blind round 3 desks A111-A113 (phase 9, added 2026-10-27)

Genuine multi-agent desks (no deliberate defect seeding) written and recorded only after the agentest v3.4 rules were frozen (sha256 recorded in agentest fixtures/standards/RULES-FREEZE-v3.4.txt). Every run was audited by hand before agentest ran. agentest column: v3.4.0, examples mode, judge on (local Qwen2.5-7B). Fails are listed as reported; the audit scores them in agentest docs/BLIND-TEST.md: 7 true positives, 13 false positives and 1 miss over the three agents.

ID	Folder	Decision	Framework	Status (7B)	Recorded runs; agentest v3.4.0
A111	a111_scholarship_award_desk	Merit scholarship award (application-intake -> merit-assessor -> disbursement-officer)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4.0 blind: 43p/4f T1-008, T1-018, T7-009, X-001
A112	a112_expense_reimbursement_desk	Travel expense reimbursement, INR/USD (expense-intake -> expense-auditor -> payments-officer)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4.0 blind: 31p/14f T1-002, T1-005, T1-006, T1-011, T1-018, T2-003, T5-001, T5-015, T7-007, T7-009, X-001, X-003, X-008, X-017
A113	a113_pharmacy_refill_desk	Pharmacy refill with interaction check (refill-intake -> pharmacist -> dispatch-clerk)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4.0 blind: 36p/2f T1-018, T7-009

Blind round 4 desks A114-A116 (phase 10, added 2026-10-28)

Genuine multi-agent desks (no deliberate defect seeding) written and recorded only after the agentest v3.4 round-4 rules freeze (sha256 in agentest fixtures/standards/RULES-FREEZE-v3.4-r4.txt). Every run was audited by hand before agentest ran. agentest column: v3.4 branch, examples mode, judge on (local Qwen2.5-7B). The audit scores the fails in agentest docs/BLIND-TEST.md (round 4): 4 true positives, 3 false positives and 1 miss (precision 0.57, recall 0.80).

ID	Folder	Decision	Framework	Status (7B)	Recorded runs; agentest v3.4
A114	a114_vendor_invoice_desk	Vendor invoice payment, 3-way match + CFO approval (ap-intake -> match-auditor -> treasury-clerk)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4 blind round 4: 40p/6f/1 not assessed T7-007, T5-015, T5-001, T2-011, T1-008, X-003
A115	a115_sim_swap_desk	Telecom SIM-swap fraud check, clean control (kyc-intake -> fraud-analyst -> provisioning-agent)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4 blind round 4: 37p/0f
A116	a116_privileged_access_desk	Privileged-access request with SoD check (access-intake -> access-reviewer -> access-provisioner)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v3.4 blind round 4: 37p/1f T7-007

Blind round 5 desks A117-A120 (phase 11, added 2026-10-28)

Genuine multi-agent desks (no deliberate defect seeding) written and recorded only after the agentest v4.0 round-5 rules freeze (agentest fixtures/standards/RULES-FREEZE-v4.0-r5.txt). Every run was audited by hand before agentest ran. agentest column: v4.0 branch, examples mode (strict identical), judge on (local Qwen2.5-7B). Models: A117/A118 Qwen2.5-7B-Instruct Q4_K_M, A119 Gemma 3 4B-it Q4_K_M, A120 Moonlight-16B-A3B-Instruct Q2_K (small open Kimi-family model, not K2); Gemma and Moonlight run behind common/gguf_generic_server.py. Scored in agentest docs/BLIND-TEST.md (round 5): 6 true positives, 3 false positives and 4 misses (precision 0.67, recall 0.60).

ID	Folder	Decision	Framework	Status	Recorded runs; agentest v4.0
A117	a117_lc_document_desk	Letter-of-credit document examination desk	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 5: 48p/1f T1-018
A118	a118_final_settlement_desk	Full-and-final settlement desk (clean control)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 5: 42p/3f T1-005, T7-009, X-008
A119	a119_card_dispute_desk	Multilingual card-dispute desk (English/Hindi/Tamil)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 5: 46p/1f T1-018
A120	a120_vendor_onboarding_desk	Vendor onboarding desk	Plain Python multi-agent + function calling (native OTel GenAI)	partial 0/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 5: 27p/4f T7-009, X-002, X-011, X-020

Blind round 6 desks A121-A126 (phase 12, added 2026-10-28)

Genuine multi-agent desks (no deliberate defect seeding) written and recorded only after the agentest v4.0 round-6 rules freeze (agentest fixtures/standards/RULES-FREEZE-v4.0-r6.txt). Every run was audited by hand before agentest ran. agentest column: v4.0 branch, examples mode (strict identical), judge on (local Qwen2.5-7B). Models: A121/A122 Qwen2.5-7B-Instruct Q4_K_M, A123/A124 Gemma 3 4B-it Q4_K_M, A125/A126 small open Kimi-family (Moonlight-16B-A3B Q2_K), not K2. Scored in agentest docs/BLIND-TEST.md (round 6): 9 true positives, 0 false positives (0 false alarms on the clean control) and 8 misses (precision 1.00, recall 0.53).

ID	Folder	Decision	Framework	Status	Recorded runs; agentest v4.0
A121	a121_records_release_desk	Medical-records release desk (privacy)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 42p/0f
A122	a122_lost_property_desk	Railway lost-property desk (clean control)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 41p/0f
A123	a123_crop_insurance_desk	Multilingual crop-insurance claim desk (English/Hindi/Telugu)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 50p/0f
A124	a124_account_takeover_desk	SaaS account-takeover response desk (security, prompt injection)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 47p/1f T7-009
A125	a125_shg_microloan_desk	Self-help-group microloan desk	Plain Python multi-agent + function calling (native OTel GenAI)	partial 0/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 20p/3f T7-009, X-002, X-011
A126	a126_power_bill_desk	Electricity bill-correction desk (English/Marathi)	Plain Python multi-agent + function calling (native OTel GenAI)	partial 0/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 6: 18p/5f T7-009, X-002, X-011, X-014, X-026

Blind round 7 desks A127-A132 (phase 13, added 2026-10-28)

Genuine multi-agent desks (no deliberate defect seeding) written and recorded only after the agentest v4.0 round-7 rules freeze (agentest fixtures/standards/RULES-FREEZE-v4.0-r7.txt). Every run was audited by hand and the audit committed before agentest ran. agentest column: v4.0 branch, examples mode (strict identical), judge on (local Qwen2.5-7B). Models: A127/A128 Qwen2.5-7B-Instruct Q4_K_M, A129/A130 Gemma 3 4B-it Q4_K_M, A131/A132 small open Kimi-family (Moonlight-16B-A3B Q2_K), not K2. Scored in agentest docs/BLIND-TEST.md (round 7): 8 true positives, 5 false positives (1 false alarm on the A127 clean control) and 11 misses (precision 0.62, recall 0.42). The "working" status is the sanity check only; the hand audit found defects in A128-A132.

ID	Folder	Decision	Framework	Status	Recorded runs; agentest v4.0
A127	a127_number_port_desk	Mobile number-portability desk (clean control)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 44p/1f T7-009 (false alarm)
A128	a128_rte_admission_desk	School RTE-quota admission desk (privacy)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 46p/2f T1-001, X-028
A129	a129_ration_grievance_desk	PDS ration-card grievance desk (English/Tamil/Bengali)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 43p/3f T7-007, X-001, X-028
A130	a130_upi_refund_desk	UPI merchant refund-dispute desk (security, prompt injection)	Plain Python multi-agent + function calling (native OTel GenAI)	working 2/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 59p/1f X-028
A131	a131_lpg_subsidy_desk	LPG subsidy (DBTL) complaint desk (English/Gujarati)	Plain Python multi-agent + function calling (native OTel GenAI)	partial 0/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 32p/3f T7-007, X-014, X-026
A132	a132_echallan_dispute_desk	Traffic e-challan dispute desk (English/Punjabi)	Plain Python multi-agent + function calling (native OTel GenAI)	partial 0/2 (proof)	+6 counterfactual runs; agentest v4.0 blind round 7: 31p/3f T7-009, X-014, X-026

Known defects

DEFECTS.md lists every real failure recorded in the proofs. D-01 to D-06 apply to the current setup: one is by design (A1), one (A4 Hindi) is a quality observation rather than a failed check, and one (A5) only appears when forced tool choice is off. D-07 to D-12 come from the 1.5B baseline runs. D-13 onwards are real failures of the complex agents A13-A25 with the 7B model. D-26 onwards come from the enterprise agents A26-A70. D-41 to D-44 come from the blind round 3 desks A111-A113 (agentest v3.4.0). D-45 to D-47 come from the blind round 4 desk A114. Each entry has its input, observed vs expected behaviour, severity, related catalogue test IDs, a reproduce command and a link to the proof.

One-command proof rerun

python verify.py --all                 # unit tests + all agents except A10, 7B model (starts/stops the model server if needed)
python verify.py --all --include-a10   # also A10 (about 10-15 min per case on CPU)
python verify.py --agent a08           # one agent
python verify.py --all --model 1.5b    # fallback model, compared with proof/baseline-1.5B/

It reruns each agent's run_demo.py (the same inputs and plain-Python checks as the proofs) into verify-out/<agent>/, never overwriting the recorded proofs. It then prints a pass/fail table, compares each case with the recorded result, and writes verify-report.json and verify-report.md. LLM output is not fully deterministic across runs, so a difference is reported, not hidden. The latest result is below.

Latest verify results (26 Oct 2026, CPU, Qwen2.5-7B-Instruct GGUF Q4_K_M):

Run	Command	Checks passed	Differences from recorded proofs
All agents except A10	`python verify.py --all`	31/34	0
A10 (run separately, 26 min)	`python verify.py --agent a10`	0/3	0
A5 with forced tool choice off	`REF_A5_UNFORCED=1 python verify.py --agent a05`	2/3	1 (expected: reproduces defect D-04)

The 3 failures in the --all run are the known defects D-01 (A8), D-02 (A7) and D-06 (A1, by design). Each run's report and outputs are in verify-runs/ (verify-report.json/.md at the root are the --all run). The checks compare pass/fail per case; the generated text differs in wording from the recorded proofs.

Proofs

Each aNN_*/proof/ folder holds run.md (a readable transcript with UTC timestamps, the model, inputs, outputs and traces) and run.jsonl (raw records). A1, A8 and A23 add n8n-execution.json; A13 onwards add trace.jsonl (the JSONL trace of every LLM call, tool call, guard decision and write). All runs were made on a Linux CPU box (8 vCPU, 15 GB RAM) on 26 Oct 2026. proof/baseline-1.5B/ holds the earlier runs with the 1.5B model.

Tests

test-mapping.csv maps each agent to catalogue test IDs (columns: agent, test_id, test_name, pack, why), with 4555 rows in total. The IDs and names are copied from the dataset CSV (catalogue/). The mapping is rule-based (scripts/build_test_mapping.py) and should be reviewed. For A1-A12 the files a*/tests/test_aNN_*.py are stubs, not real tests yet: each function only calls pytest.skip('stub: implementation coming'). A13-A25 have no stubs; instead each has real offline unit tests in tests/test_unit_aNN.py (guards, validators, mocks, workflow wiring), and common/tests/ tests the shared runtime. The mapped catalogue tests for A13-A25 are not yet automated against those agents.

Honest caveats

Default model: Qwen2.5-7B-Instruct, GGUF Q4_K_M (bartowski/Qwen2.5-7B-Instruct-GGUF), running on CPU via llama-cpp-python. It is free, needs no account, and uses about 5 GB of RAM. Hosted HF Inference Providers were not used: the free tier has only a small monthly credit, which would not cover every agent's runs.
The 7B model still makes real mistakes (A5 without forced tool choice, A7's write attempt, A8's prompt-injection leak, A10's failing code, poor Hindi in A4). These are recorded, not hidden. The 1.5B baseline runs are kept in proof/baseline-1.5B/.
Crutches: the constrained tool-name picker (1.5B server only) is not used with the 7B model. A7's schema hint was removed. A5's tool_choice='required' was kept, because without it the 7B model skipped request_refund. A named tool_choice is honoured by pre-filling the tool name, which is standard OpenAI semantics.
Human approvals in the demos are simulated by the demo scripts. The code paths that enforce them are real.
A9's 'web search' is an offline corpus. A10 runs code without Docker, using a deny-list guard and timeout, which is not a strong sandbox.
Report alternatives not used: Google ADK (A7) and Semantic Kernel (A11). A2's n8n AI-node variant is covered by A8 instead.
No third-party eval frameworks are used. Contact: srinivas AT testAIng.com

References

Pattern and framework sources (verified 2026-10-26). Cited as the origin of the pattern, not as endorsement of this implementation.

Anthropic (2024). Building effective agents (stopping conditions, workflows vs agents). https://www.anthropic.com/engineering/building-effective-agents
Yao, S., Zhao, J., Yu, D., et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629
Wang, L., Xu, W., Lan, Y., et al. (2023). Plan-and-Solve Prompting. ACL 2023. arXiv:2305.04091
Shinn, N., Cassano, F., Berman, E., et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366
Du, Y., Li, S., Torralba, A., et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.14325
Madaan, A., et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. arXiv:2303.17651
Wu, Q., Bansal, G., Zhang, J., et al. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv:2308.08155
Hong, S., et al. (2023). MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. arXiv:2308.00352
Lewis, P., Perez, E., Piktus, A., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401
Dean, J., Ghemawat, S. (2008). MapReduce: Simplified Data Processing on Large Clusters. CACM 51(1). doi:10.1145/1327452.1327492
LangGraph docs. Overview. https://docs.langchain.com/oss/python/langgraph/overview
LangGraph docs. Workflows and agents (chaining, routing, orchestrator-worker, evaluator-optimizer). https://docs.langchain.com/oss/python/langgraph/workflows-agents
CrewAI docs. Agents. https://docs.crewai.com/en/concepts/agents
CrewAI docs. Processes (sequential, hierarchical). https://docs.crewai.com/en/concepts/processes
Microsoft AutoGen docs. AgentChat user guide. https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/index.html
OpenAI Agents SDK docs. https://openai.github.io/openai-agents-python/
n8n docs. Advanced AI. https://docs.n8n.io/advanced-ai/
Model Context Protocol specification (2025-06-18). https://modelcontextprotocol.io/specification/2025-06-18
LlamaIndex docs. Introduction to RAG. https://docs.llamaindex.ai/en/stable/understanding/rag/
Qwen team. Qwen2.5-7B-Instruct model card. https://huggingface.co/Qwen/Qwen2.5-7B-Instruct
Qwen Team (2024). Qwen2.5 Technical Report. arXiv:2412.15115
OWASP Top 10 for LLM Applications 2025. LLM01:2025 Prompt Injection. https://genai.owasp.org/llmrisk/llm01-prompt-injection/
OWASP Top 10 for LLM Applications 2025. LLM06:2025 Excessive Agency. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
OWASP GenAI Security Project. OWASP Top 10 for Agentic Applications for 2026. https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

Author and affiliation

Dr Srinivas Padmanabhuni, CTO and Co-founder, AiEnsured (testAIng.com). AiEnsured is ISO/IEC 42001 certified. Contact: srinivas AT testAIng.com · testAIng.com · AiEnsured · Runner: agentest on GitHub

Citation (BibTeX)

@misc{padmanabhuni_agentic_ai_reference_agents,
  author       = {Padmanabhuni, Srinivas},
  title        = {agentic-ai-reference-agents},
  organization = {AiEnsured (testAIng.com)},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/srinivaspadmanabhuni/agentic-ai-reference-agents}},
  note         = {Agentic AI Test Checklist 110; DOI 10.5281/zenodo.23614376}
}