Agent Stack Builder
Updated
Plan the parts of an AI agent system: pick a starting point, adjust nine layers, and copy or download the result as a diagram. Stack components can be labeled with a provider. What each option means is under Definitions. To review a team's proposed agent, rebuild its stack here or open its shared link, then read the Concerns and paste the diagram into the proposal. AI agents in browsers that support WebMCP can also read the builder and fill in a stack, with your approval.
Getting started
- Pick a template. It fills the nine layers.
- Adjust the layers. Select a layer, then turn options on or off.
- Name providers (optional). Pick who supplies each part.
- Share it. Copy the diagram, a link or a review prompt.
More details
How to read the stack
- A layer is one job the system has to cover, such as Tools or Guardrails. The nine layers sit in three groups: Agent, Deployment, and Operations and maintenance.
- The number beside a layer is its position in the stack, and the blue number marks the layer that is open. Select a layer on either side to open its options. On a phone, use Add in the layer.
- Solid chips are options you have turned on, and dashed chips are options you can add. A chip with a colored outline has a provider.
- The level line under Stack, such as Level 3 of 6, shows how much autonomy the chosen options give the agent. Concerns lists gaps the builder sees in the current choices.
Custom labels and the harness
- Type your own options into the Add label field. The options are examples, and items you type in are not checked.
- The code around the model that runs the loop, passes tool calls through and enforces limits is often called the harness. Here it spans Orchestration, Guardrails and the sandbox in Runtime.
Providers and colors
- Name up to five providers, the companies or services behind the options, plus Other. Pick them from the common list or type your own.
- Assign each option with Provided by. Color marks the provider only, and the Show providers switch hides them.
The diagram
- The diagram draws the model above the runtime, also when it is self-hosted or runs on the device.
- Name the stack, then copy or download the diagram, or use Copy link to share the stack. Tap the diagram to expand it. On a phone, pinch and scroll to move around.
Templates
Examples of common stacks. Choosing one fills the stack below, and every choice can still be changed.
Stack
Level 4 of 6: Single agent
Nothing chosen yet in Model.
Providers
OtherOptional. Name up to five providers, such as a cloud or model provider.
Common providers
AgentThe model and what it works with
- 1
Reads the context and picks what to do next.
Pick a model - 2
Runs the steps or the loop.
Agent app - 3
What the system can look up and change.
Function callingMCP servers - 4
What the model is given, what is saved, and what it can look up.
InstructionsInstruction fileUser preferencesWithin a sessionCompactionAcross sessionsSaved notes
DeploymentWhere it runs and how it is reached
- 5
The machine it runs on, what starts it and how long runs recover.
Cloud computer - 6
How people and other software reach it.
Chat window
Operations and maintenanceHow it is kept safe, checked and maintained
- 7
Approvals, access, checks and limits. They apply to every layer.
Approval and reviewHuman approvalAccessPermission allowlistSecrets vault - 8
Whether it works and what it costs.
MonitoringTracingCost trackingTestingTest tasks - 9
How changes ship and who keeps it running.
Version control
Options by layer
Model
0 chosenMain
Specialized
Where it runs
Supporting
Custom
Providers
No providers named yet. Use Add custom provider or pick from common providers above.
Stack diagram
A high-level picture of the stack, useful for explaining the idea to others. Copy or download it, or copy a link to the stack, with the icons beside the heading.
Diagram of the stack named Coding agent. Model: nothing chosen. Orchestration: Agent app. Tools: Function calling, MCP servers. Memory and data: Instruction file, User preferences, Compaction, Saved notes. Runtime: Cloud computer. Interface and front end: Chat window. Guardrails: Human approval, Permission allowlist, Secrets vault. Evaluation and observability: Tracing, Test tasks, Cost tracking. Delivery and upkeep: Version control.
Definitions
Agent
The model and what it works with
1Model
Main
- Language model - Writes and reasons in text; the model most agents are built on.
- Small language model - A cheaper, faster language model for routine steps such as formatting and simple routing.
- Decision model - Returns a choice, score or yes-or-no with a probability for each allowed answer, for routing, classification and checks; it does not write text.
Specialized
- Embedding model - Turns text into vectors so a vector store can find related passages.
- Reranker - Re-orders search results by how well they answer the question.
- Speech model - Turns speech into text or text into speech.
- Image model - Reads or generates images.
Where it runs
- Hosted API - A provider runs the model and the system calls it over the network.
- Self-hosted - Model weights run on your own servers or cloud account, often an open-weight model.
- On device - The model runs on the person's laptop or phone.
Supporting
- Model gateway - One API in front of several model providers, so routing or a user setting picks which one answers.
- Fallback model - A second model, often from another provider, takes over when the main one is down, slow or refuses.
- Response cache - A repeat request reuses a saved answer instead of paying for another model call.
2Orchestration
Runs the steps
- Plain code - Fixed steps written in ordinary code, with the model called at each one.
- Agent framework - A library for building your own agent loop. Examples: LangGraph, Claude Agent SDK, OpenAI Agents SDK, Google Agent Development Kit, Microsoft Agent Framework and CrewAI.
- Agent app - An off-the-shelf product such as a coding agent or desktop assistant runs the loop. Most ask before risky actions; check the settings.
- Managed agents - A provider runs the loop and the sandbox, so there is less to operate and less to customize.
Patterns
- Subagents - A lead agent hands focused tasks to workers that start with a clean context and return a summary.
- Planning step - The agent writes a plan before it acts, then works through it.
- Self-check - The agent reviews its own answer before it replies, and fixes what it finds.
- Structured output - Model replies follow a fixed format, such as JSON, that code can check before using it.
3Tools
Connections
- Function calling - Functions written for this system that the model can call by name.
- MCP servers - Connectors to outside systems through the Model Context Protocol, reusable across agent products. Docs
Abilities
- Web search - Current information from the open web.
- Page fetching - Code retrieves a given web page and cleans it, or renders it in a headless browser, before the model reads it.
- Business APIs - Outside services the code or the model reads from or records to, such as a catalog, a metrics service or a CRM that receives leads.
- File access - Reads, writes and edits files in a folder it is given.
- Code execution - The model writes and runs code for analysis, file work and anything no tool covers.
- Computer use - The model operates a browser or desktop for systems that have no API.
- Payments - Takes or makes payments, such as charging a customer, issuing a refund or buying on someone's behalf.
4Memory and data
Instructions
- Instruction file - Standing rules loaded at the start of every session, such as an AGENTS.md file. Docs
- Skills - Packaged instructions and scripts loaded only when a task calls for them. Docs
- User preferences - What each person has set for themselves, such as tone, defaults, connected accounts and notifications.
Within a session
- Conversation history - The messages and tool results of the current session, kept so each step can see what came before.
- Compaction - A long session is summarized so work continues in a fresh context window.
Across sessions
- Saved notes - Plans, findings and shared state are written to files or a store, so the agent, or other agents, can read them later.
- Memory service - Facts and preferences are saved across sessions and recalled by the system.
- On-device storage - Data stays on the person's own device, such as in the browser or the app, instead of on a server.
- Task backlog - The list of planned work the agent picks its next task from.
Lookup
- Documents - Files, pages and knowledge bases it is allowed to search.
- Document preparation - Documents are split, indexed and embedded so they can be searched.
- Database - Structured records it can query, such as orders or accounts.
- Vector store - Documents are indexed so the relevant passages can be retrieved at run time.
- File storage - Files and media it saves or reads, such as uploads, photos, exports and generated reports.
Deployment
Where it runs and how it is reached
5Runtime
Where it runs
- Local machine - Runs on a laptop with its files and logins, and stops when the laptop does.
- Own server - The agent runs inside an application you host, on your own servers or as serverless functions, and starts with each request.
- Sandbox - An isolated container or virtual machine that limits what a mistake can reach.
- Cloud computer - A hosted machine that keeps working with no laptop open.
- Parallel sandboxes - Several agents work at once, each in its own copy of the project.
What starts it
- Schedules - Work starts on a timer, with nobody asking for it.
- Event triggers - Work starts when something happens, such as a merged change or a new message.
Long runs and recovery
- Durable workflows - Long jobs survive restarts and retry failed steps.
- Checkpoints - Progress is saved after each step, so a failed run restarts from the last good step.
- Error handling - A failed tool call is retried or reported, and the user gets a clear message when a run cannot finish.
6Interface and front end
For people
- Own app - A front end you build and host for your users.
- Chat window - A chat screen in an app or on a web page.
- Browser extension - It runs inside the person's browser and works on the page they have open.
- Team chat - People message it, and it posts results, in a chat workspace the team already uses.
- Email - It reads requests from an inbox and replies or sends reports by email.
- In another app - A button or side panel in a tool people already work in, such as a document editor or a help desk.
- Command line - A person runs it from a terminal, or a script starts it.
- Voice or phone - Requests are spoken and replies are read aloud, in an app or on a phone line.
- File exports - Results leave as files, such as CSV, spreadsheets, PDF or Markdown.
For software and agents
- API - Other software sends it requests and reads its results.
- Agent endpoint - Other agents can call it as a tool, for example through an MCP server it offers.
Front-end features
- Streaming replies - The answer appears word by word while it is still being written.
- Steps and sources - The answer shows the steps taken and the sources used.
- Image and file upload - People send photos, screenshots or documents for the model to read.
Operations and maintenance
How it is kept safe, checked and maintained
7Guardrails
Approval and review
- Human approval - A person confirms sends, purchases, deletions and production changes.
- Human handoff - The system passes the task to a person when it cannot finish or is unsure.
- Review model - Another model, or a panel of models, reviews output and checks each risky action before it runs.
Access
- Permission allowlist - A written list of the actions that may run with no confirmation.
- Secrets vault - Credentials are supplied at call time so they stay out of prompts and logs, where the product supports it.
- Sign-in and roles - Each person signs in, and the agent acts with that person's access or a named service account.
- Abuse protection - A firewall, bot check and per-visitor rate limits keep attacks, scripts and floods from reaching it or running up cost.
- Network policy - A list of the hosts the runtime may reach.
- Privacy and consent - People agree before their details are kept, and personal data is held only as long as needed.
Automatic checks
- Content filters - Checks on what comes in and what goes out, such as moderation, personal-data redaction and prompt-injection screening.
- Code and rule checks - A program checks every change against rules, such as types, lint, formats and architecture rules.
Limits and stopping
- Spend limit - A cap per run and per day that stops work when it is reached.
- Stop switch - One control halts every run at once when something goes wrong.
- Reversible actions - Changes are made as drafts, soft deletes or logged edits, so a bad one can be undone.
8Evaluation and observability
Monitoring
- Tracing - A record of every model call and tool call in a run.
- Audit log - A record of who approved what and which actions ran.
- Cost tracking - Spend per task and per agent.
- Uptime monitoring - Tracks how long runs take, how often they fail, whether the service is up and how much compute it uses.
- Usage analytics - Shows who uses it, how often and for which tasks, so effort goes where the use is.
- Alerts - Someone is told when runs start failing or spend jumps.
Testing
- Test tasks - Saved tasks with known good outcomes, rerun after every change.
- Proof of work - Screenshots, recordings or test results that show a change works.
- User feedback - Thumbs and flags from people, corrections and production signals feed the next round of work.
9Delivery and upkeep
Changes
- Version control - Prompts, tool definitions and settings are kept in a repository with their history.
- Model pinning - A fixed model version is named in the settings, with a test run planned before moving to a newer one.
Releases
- Release tests - The test tasks run automatically before a change reaches users.
- Deploy on merge - A merged change builds and goes live on its own, with a preview for each branch.
- Staging - A separate copy of the system for trying a change before users see it.
- Rollback - A release is a recorded step, and a bad one can be undone quickly.
Upkeep
- Runbook and owner - A named person keeps it running, with written steps for common failures.
- Scheduled updates - Model, SDK and tool versions are reviewed and updated on a set schedule.
Suggested next steps
These steps turn the stack into a plan that other people can review and act on. Several suit someone who approves or funds the work and will not build it, such as naming owners, raising risk questions and sharing the diagram.
- Map the workflow. Write out each step, branch and approval, then draw it with Mermaid, as in the Whisk architecture diagram.
- Name an owner for each layer. Say who sets it up and who runs it, such as IT for the runtime and security for guardrails.
- Raise data and risk questions early. Ask security, legal or compliance about the data the agent reads and the actions it takes before anything is built.
- Agree on a first task. Pick one task with a clear result, a small group of users and a date to review how it went.
- Share the diagram. Put it in the proposal so reviewers see the same picture, and list any empty layers as open questions.
- Write a shared glossary. List the domain terms the agent and the team both use, with one meaning for each, and load it as a standing instruction.
- Have an expert review the answers. Ask someone who knows the subject to read a sample of the agent's answers and mark what is wrong.
- Talk it through with an AI assistant. Copy the prompt below into the assistant you use. It can find gaps, question the choices and help plan the first build, or list questions to ask the team that will build it.
Prompt
Review this agent stack. It lists the parts of one agent system in nine layers, grouped as Agent, Deployment, and Operations and maintenance. Tell me what looks missing, risky or more than the job needs, layer by layer, then name the three things to set up first. Ask me about the purpose first if it is unclear. If I am reviewing a plan someone else will build, also list the questions I should ask that team. Made with the Agent Stack Builder: https://ryansmedstad.com/projects/agent-stack-builder/?v=2&for=group&stack=agent-app.function-calling.mcp.instruction-file.user-settings.notes.cloud-agent.chat-window.approval.allowlist.vault.tracing.cost-tracking.eval-set.version-control.compaction&name=Coding+agent Definitions of every option: https://ryansmedstad.com/projects/agent-stack-builder/#stack-definitions Builder version v11.19.2026. Stack name: Coding agent Autonomy level: 4 of 6, Single agent 1. Model (Agent) - Nothing chosen 2. Orchestration (Agent) - Agent app: An off-the-shelf product such as a coding agent or desktop assistant runs the loop. Most ask before risky actions; check the settings. 3. Tools (Agent) - Function calling: Functions written for this system that the model can call by name. - MCP servers: Connectors to outside systems through the Model Context Protocol, reusable across agent products. 4. Memory and data (Agent) - Instruction file: Standing rules loaded at the start of every session, such as an AGENTS.md file. - User preferences: What each person has set for themselves, such as tone, defaults, connected accounts and notifications. - Compaction: A long session is summarized so work continues in a fresh context window. - Saved notes: Plans, findings and shared state are written to files or a store, so the agent, or other agents, can read them later. 5. Runtime (Deployment) - Cloud computer: A hosted machine that keeps working with no laptop open. 6. Interface and front end (Deployment) - Chat window: A chat screen in an app or on a web page. 7. Guardrails (Operations and maintenance) - Human approval: A person confirms sends, purchases, deletions and production changes. - Permission allowlist: A written list of the actions that may run with no confirmation. - Secrets vault: Credentials are supplied at call time so they stay out of prompts and logs, where the product supports it. 8. Evaluation and observability (Operations and maintenance) - Tracing: A record of every model call and tool call in a run. - Test tasks: Saved tasks with known good outcomes, rerun after every change. - Cost tracking: Spend per task and per agent. 9. Delivery and upkeep (Operations and maintenance) - Version control: Prompts, tool definitions and settings are kept in a repository with their history. How the parts connect: - A request comes in through the interface and reaches orchestration. - Orchestration, tools and memory run inside the runtime. - No model is chosen yet. - No subagents are included. - Guardrails apply to every layer. - Evaluation and observability watch the whole system, and delivery and upkeep ship and maintain it. - Layers with nothing chosen: Model.
You can edit this prompt. Edits stay on this page.
Further Reading
- Building an Agentic System, Layer by Layer
Starts with workflows and agents, where fixed code or the model picks each step, then walks through the layers, the choices in each and what can go wrong. It ends with a working example, an order of setup for builders and questions for leaders to ask before approving a build.
MCP integration for AI agents
WebMCP is a browser feature that lets a page offer tools to an AI agent in the same tab. In a browser that supports it, this builder offers four tools. They work only inside this page, load nothing from other sites, save nothing, send nothing and use no personal data.
list_stack_options- Lists the nine layers, or the options in one layer with their ids and definitions.
get_current_stack- Returns the stack name, the providers, the options chosen in each layer with the provider of each, empty layers, the concerns the builder flagged and a link to the stack. Names the visitor typed come back as data to read.
get_review_prompt- Returns the review prompt the page builds from the current stack.
set_stack- Asks to choose options by id, with an optional name, providers and a provider for each option. Nothing changes until the visitor approves, and Undo reverses it. It takes only known option ids, never custom options. Replace is the default, and add keeps the current options.
- Agents without WebMCP
-
Request this page with
Accept: text/markdownfor a Markdown copy with every option id. To hand a visitor a stack, build a link like this one, which opens a multi-agent research helper:/projects/agent-stack-builder/?v=2&stack=llm.hosted-api.agent-framework.subagents&prov=Anthropic~purple&on=llm~0&name=Research+helper. Ids the builder does not know are ignored.
Option ids by layer
Stack link format: option ids are joined by dots in stack. Each prov value is a provider name and a color joined by ~. Each on value pairs an option id with a provider number, counted from 0, or o for Other, joined by ~, and the pairs are joined by dots.
Builder version v11.19.2026. Provider colors: orange, blue, purple, green, pink.
Model
-
llm: Language model -
small-model: Small language model -
decision-model: Decision model -
embedding-model: Embedding model -
reranker: Reranker -
speech-model: Speech model -
image-model: Image model -
hosted-api: Hosted API -
self-hosted: Self-hosted -
on-device-model: On device -
model-gateway: Model gateway -
fallback-model: Fallback model -
response-cache: Response cache
Orchestration
-
plain-code: Plain code -
agent-framework: Agent framework -
agent-app: Agent app -
managed-agents: Managed agents -
subagents: Subagents -
structured-output: Structured output -
planning: Planning step -
self-check: Self-check
Tools
-
function-calling: Function calling -
mcp: MCP servers -
data-apis: Business APIs -
page-fetch: Page fetching -
web-search: Web search -
payments: Payments -
code-execution: Code execution -
browser: Computer use -
file-access: File access
Memory and data
-
skills: Skills -
instruction-file: Instruction file -
user-settings: User preferences -
documents: Documents -
file-storage: File storage -
database: Database -
compaction: Compaction -
notes: Saved notes -
vector-store: Vector store -
history: Conversation history -
on-device: On-device storage -
memory-service: Memory service -
doc-prep: Document preparation -
backlog: Task backlog
Runtime
-
local: Local machine -
server: Own server -
container: Sandbox -
cloud-agent: Cloud computer -
schedules: Schedules -
event-triggers: Event triggers -
parallel-sandboxes: Parallel sandboxes -
durable-workflow: Durable workflows -
checkpoints: Checkpoints -
error-handling: Error handling
Interface and front end
-
own-app: Own app -
chat-window: Chat window -
chat-channel: Team chat -
email: Email -
in-app: In another app -
command-line: Command line -
file-exports: File exports -
browser-extension: Browser extension -
api: API -
agent-endpoint: Agent endpoint -
voice: Voice or phone -
image-input: Image and file upload -
streaming: Streaming replies -
show-steps: Steps and sources
Guardrails
-
approval: Human approval -
allowlist: Permission allowlist -
network-policy: Network policy -
vault: Secrets vault -
sso: Sign-in and roles -
review-model: Review model -
human-handoff: Human handoff -
stop-switch: Stop switch -
reversible-actions: Reversible actions -
bot-check: Abuse protection -
privacy: Privacy and consent -
filters: Content filters -
automated-checks: Code and rule checks -
spend-limit: Spend limit
Evaluation and observability
-
tracing: Tracing -
eval-set: Test tasks -
cost-tracking: Cost tracking -
audit-log: Audit log -
alerts: Alerts -
feedback: User feedback -
proof-of-work: Proof of work -
performance: Uptime monitoring -
usage-analytics: Usage analytics
Delivery and upkeep
-
version-control: Version control -
ci-tests: Release tests -
model-pinning: Model pinning -
deploy-on-merge: Deploy on merge -
staging: Staging -
rollback: Rollback -
runbook: Runbook and owner -
scheduled-updates: Scheduled updates
Sources
- Anthropic - Building effective agents
- Anthropic - Effective context engineering for AI agents
- OpenAI - A practical guide to building agents (PDF)
- LangChain - How JP Morgan Built An AI Agent for Investment Research with LangGraph
- AI Engineer - Software Fundamentals Matter More Than Ever (Matt Pocock)
- AI Engineer - All Code Is Slop (Except Mine) (Lauren Tan)
v11.19.2026