Answer once and get one pick per layer of an AI agent product, from 583 catalogued systems. Every reason is a
catalogued fact with its source. Picks are ranked on fit: your must-haves, documented integrations with earlier picks,
maturity and how well the catalog documents the system; GitHub stars only break ties. Archived and
shut-down systems are never picked. A pair with no recorded integration is not recorded, not incompatible.
Reference stacks for each product type, checked against public production write-ups, are in the blueprints.
Risklow(1.5)
Integrations documented21 of 36 pairs
OWASP risks addressed8 of 10
Agent framework
Defines agents, tool calls and control flow in code.
CrewAI: Tied on fit (score 7.5); LangChain (framework) ranks first on the tiebreaker, adoption (149000 vs 58600 GitHub stars)
LlamaIndex (framework): Tied on fit (score 7.5); LangChain (framework) ranks first on the tiebreaker, adoption (149000 vs 53000 GitHub stars)
Exit path: self-host yes, open source yes; fallback CrewAI. Switching: no published price for one side; no documented integration with the other picks.
Deployment: Managed-only — OpenAI API + ChatGPT, also Microsoft Azure and AWS Bedrock (separate gpt-oss open-weight models exist)↗
Pricing: API flagship GPT-6 Astra $10 input / $50 output per 1M tokens (Fast mode 2x price); ChatGPT Free / Go / Plus / Pro / Business / Enterprise plans↗
Recently released: GPT-6.1 Terra (2026-10-26; computer use) and GPT-6.1 Sol; earlier GPT-6 Sol / Luna / Instant in ChatGPT (2026-10-07) and API flagship GPT-6 Astra (announced 2026-09-03)↗
Why not Anthropic Claude (foundation models), Google Gemini 3 family?
Anthropic Claude (foundation models): No recorded integration with LangChain (framework); Score 11 vs 11.3 for OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4)
Google Gemini 3 family: No recorded integration with LangChain (framework); Score 11 vs 11.3 for OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4)
Exit path: self-host no, open source no; fallback Google Gemini 3 family. Switching: Google Gemini 3 family saves about $7,200 a month at this workload, from published prices; works with LangChain (framework) already in this stack.
Model gateway
One API in front of every model: routing, fallbacks, caching, budgets and rate limits.
Works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4): LiteLLM docs have an OpenAI provider page.↗
Works with LangChain (framework): LangChain docs have a LiteLLM integration provider page (Unified interface for 100+ LLM APIs).↗
Licence: MIT (except enterprise/ directory, commercial license)↗
Deployment: Self-hosted only — SDK and proxy; Enterprise self-hosted or air-gapped↗
MCP: yes — MCP Gateway in LiteLLM Proxy (Streamable HTTP, SSE, stdio) with per-key/team access control↗
Why not OpenRouter, TrueFoundry AI Gateway?
OpenRouter: No recorded integration with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no licence, deployment, mcp fact for it; Score 8.5 vs 10.5 for LiteLLM
TrueFoundry AI Gateway: No recorded integration with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no licence, deployment, mcp fact for it; Score 6.8 vs 10.5 for LiteLLM
Exit path: self-host yes, open source yes; fallback agentgateway. Switching: no published price for one side; works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) already in this stack.
Memory layer
Remembers users, facts and past sessions across conversations.
Works with LiteLLM: Mem0 docs have a LiteLLM LLM provider page for Mem0 open source.↗
Works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4): Mem0 docs have an OpenAI as LLM Provider LLM provider page for Mem0 open source.↗
Works with LangChain (framework): Mem0 docs have a LangChain Integration integration page (Build personalized AI agents using LangChain for conversation flow and Mem0 for long-term memory retention).↗
Exit path: self-host yes, open source yes; fallback Zep & Graphiti. Switching: no published price for one side; works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) already in this stack.
Vector store
Stores embeddings for retrieval (RAG) and semantic search.
Weaviate: No recorded integration with Mem0, LiteLLM, LangChain (framework); Tied on fit (score 12); Milvus ranks first on the tiebreaker, adoption (46000 vs 16800 GitHub stars)
Elasticsearch: No recorded integration with Mem0, LiteLLM, LangChain (framework); The catalog has no sdks, recently released fact for it; Score 10.6 vs 12 for Milvus
Marqo: Ruled out: deprecated — superseded by a successor product
Exit path: self-host yes, open source yes; fallback Weaviate. Switching: Milvus has no published price; Weaviate would cost $45 a month at this workload; works with LangChain (framework), OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), Mem0 already in this stack.
Embeddings / reranker
Turns text into vectors and reranks retrieved results.
Cohere Embed: No recorded integration with Milvus, Mem0, OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); Score 10.1 vs 12.9 for OpenAI Embeddings
Sentence Transformers: No recorded integration with Milvus, Mem0, OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no recently released fact for it; Score 10 vs 12.9 for OpenAI Embeddings
Exit path: self-host no, open source no; fallback Cohere Embed. Switching: Cohere Embed has no published price (self-hosted: your own infrastructure); works with LangChain (framework), Milvus already in this stack.
Tools, auth and MCP
Gives agents authenticated access to external apps and APIs, scoped per user.
Works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4): Composio's SDKs and frameworks page lists a OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) provider.↗
Works with LangChain (framework): Composio's SDKs and frameworks page lists a LangChain (framework) provider.↗
Deployment: Managed-only — hosted API + MCP endpoints (Helm charts published for Kubernetes deploys)↗
Why not Arcade.dev, IBM ContextForge MCP Gateway?
Arcade.dev: No recorded integration with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); Score 6.4 vs 8.9 for Composio
IBM ContextForge MCP Gateway: No recorded integration with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no sdks, deployment, mcp fact for it; Score 6.2 vs 8.9 for Composio
Exit path: self-host yes, open source yes; fallback Arcade.dev. Switching: Arcade.dev has no published price (self-hosted: your own infrastructure); works with LangChain (framework) already in this stack.
Evals & observability
Traces every step, runs evaluations, catches regressions.
Works with Milvus: Milvus's integrations overview lists a Milvus + Arize Phoenix integration tutorial.↗
Works with LiteLLM: Arize Phoenix docs have a LiteLLM integration page.↗
Works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4): Arize Phoenix docs have a OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) integration page.↗
Works with LangChain (framework): Arize Phoenix docs have a LangChain (framework) integration page.↗
MCP: yes — built-in remote MCP server at /mcp; legacy stdio @arizeai/phoenix-mcp in maintenance↗
Why not OpenLLMetry (Traceloop), Braintrust, Humanloop, Patronus AI?
OpenLLMetry (Traceloop): No recorded integration with Milvus, LiteLLM, OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no mcp fact for it; Score 12.5 vs 13.3 for Arize Phoenix
Braintrust: No recorded integration with Milvus, LiteLLM, OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LangChain (framework); The catalog has no adoption (tiebreaker) fact for it; Score 11.8 vs 13.3 for Arize Phoenix
Exit path: self-host unknown, open source no; fallback OpenLLMetry (Traceloop). Switching: OpenLLMetry (Traceloop) is beyond its published plans at this workload (contact sales); works with LangChain (framework), OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LiteLLM already in this stack.
Guardrails & security
Screens inputs and outputs for prompt injection, data leaks and unsafe actions.
Presidio: No recorded integration with OpenAI Embeddings, LangChain (framework); Score 7 vs 7.7 for NVIDIA NeMo Guardrails
Google Cloud Model Armor: No recorded integration with OpenAI Embeddings, LangChain (framework); The catalog has no licence, recently released, adoption (tiebreaker) fact for it; Score 5.8 vs 7.7 for NVIDIA NeMo Guardrails
Exit path: self-host yes, open source yes; fallback Presidio. Switching: no published price for one side; works with OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4), LiteLLM already in this stack.
Do the picks work together?
LangChain (framework) + OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) — LangChain docs have an OpenAI integration provider page (GPT models and comprehensive AI platform).↗
LangChain (framework) + LiteLLM — LangChain docs have a LiteLLM integration provider page (Unified interface for 100+ LLM APIs).↗
LangChain (framework) + Mem0 — Mem0 docs have a LangChain Integration integration page (Build personalized AI agents using LangChain for conversation flow and Mem0 for long-term memory retention).↗
LangChain (framework) + Milvus — LangChain docs have a Milvus integration provider page (Open-source vector database for AI applications).↗
LangChain (framework) + OpenAI Embeddings — LangChain docs have an OpenAI embeddings integration page (OpenAIEmbeddings).↗
LangChain (framework) + Composio — Composio's SDKs and frameworks page lists a LangChain (framework) provider.↗
LangChain (framework) + Arize Phoenix — Arize Phoenix docs have a LangChain (framework) integration page.↗
LangChain (framework) + NVIDIA NeMo Guardrails — NeMo Guardrails docs have a LangChain integration page.↗
OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) + LiteLLM — LiteLLM docs have an OpenAI provider page.↗
OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) + Mem0 — Mem0 docs have an OpenAI as LLM Provider LLM provider page for Mem0 open source.↗
Only cited facts add points: lifecycle status, several layers from one vendor, low contributor count (OpenDigger), source-available or changed licences, integrations the catalog can't confirm, and layers with no option.
+0.5 NVIDIA NeMo Guardrails: beta, not generally available↗
+1 2 layers from one vendor (openai): Foundation model, Embeddings / reranker
Security coverage — OWASP Top 10 for LLM applications
Addressed means the stack has a component aimed at the risk, not that it's mitigated. OWASP list ↗
Risk
Status
By
What to do
LLM01 Prompt injection
addressed
NVIDIA NeMo Guardrails
Screen user input and retrieved content before it reaches the model.
LLM02 Sensitive information disclosure
addressed
NVIDIA NeMo Guardrails
Detect and redact PII and secrets in prompts and outputs.
LLM03 Supply chain
practice
Run the stack check (GitHub Action) on every PR so archived or acquired components are flagged.
LLM04 Data and model poisoning
addressed
Arize Phoenix
Evaluate on fixed test sets and watch for drift after data or model changes.
LLM05 Improper output handling
addressed
NVIDIA NeMo Guardrails
Validate model output before it reaches other systems; run generated code in a sandbox.
LLM06 Excessive agency
addressed
Composio
Give agents scoped, per-user credentials and isolate what they execute.
LLM07 System prompt leakage
addressed
NVIDIA NeMo Guardrails
Keep secrets out of prompts; filter outputs that echo them.
LLM08 Vector and embedding weaknesses
practice
Enforce per-tenant access control in the vector store and memory layer; never mix tenants in one index without filters.
LLM09 Misinformation
addressed
Arize Phoenix
Evaluate groundedness and accuracy before and after release.
LLM10 Unbounded consumption
addressed
LiteLLM
Rate limits, budgets and timeouts on model calls.
Monthly cost: $9,200 to $9,538plus 1 "contact sales" line
$9,030 from published list prices (checked 2026-10-27), the rest self-hosting estimates for components with no published price (assumptions below, checked 2026-10-31). Usage beyond the largest published plan is "contact sales", never extrapolated.
Layer
System
Monthly
Plan / working
Agent framework
LangChain (framework)
$0.00 to $0.00 (est.)
Self-hosting estimate: A library inside your application; no separate infrastructure..
gpt-6-astra list price 180,000 model calls a month · 540.0M input tokens × $10.00/M + 72.0M output tokens × $50.00/M = $9,000 · With gpt-6-luna instead: $90.00 a month Before prompt caching or batch discounts.
Model gateway
LiteLLM
$30.37 to $70.08 (est.)
Self-hosting estimate: One proxy instance.. AWS t3.medium (2 vCPU, 4 GiB) at $0.04/h to AWS m6i.large (2 vCPU, 8 GiB) at $0.10/h, 730 h a month.↗1↗2
Enterprise (custom); no overage rates published 60,000 memory writes and 60,000 searches a month Usage is beyond the largest published plan; the vendor quotes this privately.
Vector store
Milvus
$70.08 to $184 (est.)
Self-hosting estimate: One node; index size drives memory.. AWS m6i.large (2 vCPU, 8 GiB) at $0.10/h to AWS r6i.xlarge (4 vCPU, 32 GiB) at $0.25/h, 730 h a month.↗1↗2
text-embedding-3-small list price 30.0M tokens embedded a month × $0.02/M = $0.60 Ongoing embedding only; embedding an existing corpus once is extra. text-embedding-3-large is $0.13 per 1M tokens (3,072 dimensions).
Pro 120,000 tool calls a month · Pro: $29.00 includes 196,666 Pro includes the 100k free calls plus a $29 usage credit, converted at the $0.0003 base rate (196,666 calls); assumes your own OAuth apps (Composio-managed apps get 20k free, then +$0.0002/call); the free Hobby plan (100k calls) is hard-capped, so it is left out.
Evals & observability
Arize Phoenix
$70.08 to $184 (est.)
Self-hosting estimate: One collector instance plus its database.. AWS m6i.large (2 vCPU, 8 GiB) at $0.10/h to AWS r6i.xlarge (4 vCPU, 32 GiB) at $0.25/h, 730 h a month.↗1↗2
Guardrails & security
NVIDIA NeMo Guardrails
$0.00 to $70.08 (est.)
Self-hosting estimate: A library in-process, or one service instance.. AWS m6i.large (2 vCPU, 8 GiB) at $0.10/h, 730 h a month.↗1
Infrastructure estimates for components with no published price, assumed self-hosted on one reference instance (monthly = hourly list price × 730 h). AWS hourly prices are read from EC2Instances.info (Vantage), a third-party price index, and are estimates; the Lambda price is the vendor's own page. Storage, egress, redundancy and operations time are not included.
Security review sheet
For procurement and security review: attestations, data handling, encryption, SSO/RBAC, deployment, licence, SDKs, protocols, maturity, HQ and pricing for every pick, each with the page it was read from.
80 of 126 values are published; the rest read "not published" and link the vendor's trust page where the catalog has one.
Save this next to your code. npx landscape check stack.yml, the stack check and the "AI stack check" GitHub Action flag any component that is later archived, acquired or goes stale.
# Agent stack — generated by the AI Agent Infrastructure Landscape (plan_stack).
# Check it any time: npx landscape check stack.yml, or the "AI stack check" GitHub Action.
version: 1
product: support-agent
requirements:
language: any
hosting: any
oss: false
mcp: false
components:
framework: langchain-framework--langchain-com # LangChain (framework)
model: openai-gpt-family-gpt-5-gpt-4o-o3-o4--openai-com # OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4)
gateway: litellm--gh-berriai-litellm # LiteLLM
memory: mem0--mem0-ai # Mem0
vectors: milvus--milvus-io # Milvus
embeddings: openai-embeddings--platform-openai-com # OpenAI Embeddings
tools: composio--composio-dev # Composio
observability: arize-phoenix--phoenix-arize-com # Arize Phoenix
guardrails: nvidia-nemo-guardrails--docs-nvidia-com # NVIDIA NeMo Guardrails