Stop Token :: stack map
217 tools across 15 layers in 6 blocks; 109 of them open source. The layers are the fifteen of the article, numbered and ordered the same way. Tiers are our reading of each vendor’s model, not a quote. Open core means free to run yourself and metered in the vendor’s cloud. Price bands are the tool’s own price per month in dollars, for a first deployment of 5–50 users, not the cost of building the layer; the chooser reports a range of 0.6 to 1.5 times the band’s middle.
Human view: /stack
## 01 Model (28)
Writes the answer: a hosted API, or open weights on an inference server you run.
Stage 1 of 3 in the article’s order.
Anthropic API — first choice
OpenAI API — first choice
Amazon Bedrock models
Azure OpenAI
Cortex AI models
Gemini API
Alibaba Cloud Model Studio
Cohere
Databricks Model Serving
DeepSeek API
Fireworks
Groq
Mistral
Modal
Together
Baseten
vLLM — first choice
Ollama
SGLang
Qwen (Alibaba) — first choice
DeepSeek
GLM (Zhipu / Z.ai)
Kimi K2 (Moonshot)
gpt-oss (OpenAI)
Gemma (Google)
Llama (Meta)
MiniMax M-series
Mistral open weights
## 02 System prompt (2)
The role and the limits, versioned and tested like code.
Stage 1 of 3 in the article’s order.
Prompts in git
PromptLayer
## 03 Knowledge base (32)
Documents, the data catalogue and the links between them, searched within the user's rights.
Stage 1 of 3 in the article’s order.
DataHub
Atlan
Dataplex
Microsoft Purview
Unity Catalog
LlamaIndex — first choice
Haystack
Neo4j — first choice
Cognee
LightRAG
Microsoft GraphRAG
Docling — first choice
Azure Document Intelligence
Unstructured
Amazon Textract
Document AI
LlamaParse
pgvector — first choice
Cortex Search
Databricks Vector Search
Qdrant
Vertex AI Search
Azure AI Search
Bedrock Knowledge Bases
BigQuery
Chroma
Elasticsearch
LanceDB
OpenSearch
Pinecone
Weaviate
Turbopuffer
## 04 Semantic layer (14)
One definition of each metric and entity, and questions answered over it rather than over raw tables.
Stage 3 of 3 in the article’s order.
Cube — first choice
Fabric and Power BI semantic models
Snowflake Semantic Views
Unity Catalog metric views
dbt Semantic Layer
Looker (LookML)
Power BI semantic model
AI/BI Genie
Cortex Analyst
Wren AI
Amazon Q in QuickSight
Copilot in Power BI
Gemini in Looker
Vanna
## 05 Data quality (7)
Says whether the tables behind an answer can be trusted today.
Stage 3 of 3 in the article’s order.
Soda
AWS Glue Data Quality
Elementary
Great Expectations
Horizon data metric functions
Monte Carlo
Lakehouse Monitoring
## 06 Harness (32)
Runs the loop: the plan, the calls, the limits, the point where a person signs off.
Stage 1 of 3 in the article’s order.
LangGraph — first choice
Microsoft Agent Framework
OpenAI Agents SDK
Pydantic AI
CrewAI
Google ADK
Goose
Mastra
Strands Agents
Agno
Hermes Agent
LangChain
OpenClaw
Microsoft AutoGen
Semantic Kernel
Microsoft Foundry — first choice
Amazon Bedrock AgentCore
Databricks Agent Bricks
Gemini Enterprise Agent Platform
Snowflake Cortex Agents
Salesforce Agentforce
Agyn
Temporal — first choice
Hatchet
Inngest
Restate
Trigger.dev
AWS Lambda
DBOS
Dify
n8n
Langflow
## 07 Memory (3)
Remembers the conversation and the customer between sessions.
Stage 2 of 3 in the article’s order.
Mem0
Letta Code
Zep / Graphiti
## 08 MCP & skills hub (31)
Tools, skills and sandboxes from one catalogue: what the agent can do, and where its code runs.
Stage 1 of 3 in the article’s order.
Databricks managed MCP servers
MCP Toolbox for Databases
Snowflake
dbt
ClickHouse
PostgreSQL
MCP — first choice
A2A
MCP Registry
Archestra.ai
Docker MCP Toolkit
MetaMCP
Smithery
AgentCore Code Interpreter
Azure Container Apps dynamic sessions
E2B
Daytona
AWS MCP servers
Azure MCP Server
GitHub
Playwright
Composio
GitLab
Linear
Notion
Sentry
Slack
Stripe
Vault
Nextcloud
Agent Skills — first choice
## 09 Channel + handoff (8)
Where people talk to it, and how a hard case reaches a person.
Stage 1 of 3 in the article’s order.
Intercom Fin
Zendesk AI agents
LibreChat — first choice
Microsoft Copilot Studio
Open WebUI
Amazon Q Business
Gemini Enterprise
Snowflake Intelligence
## 10 Identity + access (18)
Acts on the user's behalf or under its own identity, with the right permissions and secrets.
Stage 2 of 3 in the article’s order.
AgentCore Identity
Microsoft Entra Agent ID
Okta for AI Agents
Auth0 for AI Agents
Descope
WorkOS
OpenFGA
Cerbos
OPA
OpenBao — first choice
AWS Secrets Manager
Azure Key Vault
Infisical
Secret Manager
Databricks secrets
Doppler
HashiCorp Vault
AWS IAM
## 11 Filter (13)
Catches prompt injection on the way in and checks answers on the way out.
Stage 1 of 3 in the article’s order.
Azure AI Content Safety
Bedrock Guardrails
Lakera (Check Point)
Model Armor
NeMo Guardrails
Cisco AI Defense
Cortex Guard
LLM Guard
Palo Alto Prisma AIRS
HiddenLayer
Zenity
Guardrails AI
Llama Guard
## 12 Gateway (12)
One door for every model call: masks personal data before it leaves, keys and budgets, a fallback when a provider fails, a cache that respects rights, routing by price and quality, a log of every call.
Stage 1 of 3 in the article’s order.
Cloudflare AI Gateway
Presidio
AI Gateway guardrails
LiteLLM — first choice
AgentCore Gateway
Azure API Management AI gateway
Mosaic AI Gateway
Portkey
Apigee
Kong AI Gateway
OpenRouter
RouteLLM
## 13 Traces (7)
Records every step the system took.
Stage 1 of 3 in the article’s order.
Langfuse — first choice
Arize Phoenix
LangSmith
MLflow
Opik
W&B Weave
CloudWatch
## 14 Evals (9)
Tests every change against a set of reference questions.
Stage 1 of 3 in the article’s order.
Promptfoo
DeepEval
Bedrock Evaluations and CloudWatch
Foundry evaluations and observability
MLflow agent evaluation
Vertex AI Evaluation
Braintrust
Cortex AI Observability
Ragas
## 15 Feedback (1)
Turns bad answers from real users into tests and fixes.
Stage 3 of 3 in the article’s order.