Tools Free, no sign-up, runs in your browser Model data checked

Agent Architecture Playground

Sketch a multi-agent system on a canvas. Get a cost per run and a rule-based review of loops, unapproved actions and prompt-injection paths.

Loading Agent Architecture Playground…

What it is for

Most multi-agent designs get drawn on a whiteboard, approved, and built before anyone asks what they cost or what an injected instruction could do with them. The playground is a whiteboard that asks those questions. Draw the agents, tools, data stores and approval steps, connect them, and the panel on the right updates with a cost per run and a design review.

It models a design. It does not run agents, call tools or execute anything you type.

The building blocks

  • Input: where work enters. Mark it untrusted if a third party can write it: public users, inbound email, web pages.
  • Agent: a model with a role, calls per run, token sizes, an optional iteration cap and an optional dollar budget.
  • Tool: read-only, write, or irreversible (money, deletes, deploys, messages). Flag whether it sends data outside the company, touches private data, or returns untrusted content.
  • Human approval: a person who must approve before anything downstream happens.
  • Data store: memory, a vector store, a database. Flag it if it holds private data.
  • Output: where results land. Flag it if it leaves the company.

An arrow from an agent to a tool means “can call”. An arrow between agents means work is handed over. Arrows back and forth form a loop, and a loop needs an iteration cap.

What the review checks

The review is a fixed set of rules, the same every time, and it shows which nodes each finding comes from.

  • Prompt-injection exfiltration path (critical): one agent reads untrusted content, can see private data, and can send things out, with no human in between. That combination is how most real prompt-injection incidents work.
  • Irreversible action without approval (critical) and write access without approval.
  • Untrusted input can drive writes: content a third party wrote reaches an agent with write tools.
  • Loop with no iteration cap, no spend cap and estimate exceeds budget.
  • Long agent chain: if each step is right 95% of the time, five in a row are right 77% of the time.
  • Context larger than the model window, model without tool calling wired to tools, wide fan-out, public output with no review, single model provider, unconnected nodes.

No findings does not mean the design is safe, only that none of these rules fired. For the threat model behind them, read Threat Modeling an AI Agent.

Cost per run

Each agent costs calls × (input tokens × input price + output tokens × output price). Agents inside a loop are multiplied by the loop’s iteration cap, or by 3 if there is none (and the review flags the missing cap). That figure deliberately ignores context growth and retries. Open this design in the cost simulator carries the agents over so you can model those properly.

Templates

Five starting points: a single tool-using agent, a Researcher → Analyst → Writer pipeline with editorial review, a Planner → Developer → Tester loop that ends in a PR review, a supervisor with three workers sharing memory, and an “autonomous organization” that deliberately trips most of the rules. Load the last one to see what the review catches.

Related

FAQ

What is a multi-agent architecture?

A system where several LLM-driven agents, each with its own role, prompt and tools, pass work between them: a planner and workers, a pipeline of specialists, or a supervisor that delegates. It can help with complex tasks. It also multiplies cost, latency and the number of places things go wrong, which is what the review is for.

Can I export my diagram?

Yes: JSON (re-importable), SVG and PNG images, and Mermaid text that renders in GitHub, GitLab, Notion and most Markdown tools. Copy share link puts the whole design in the URL.

Does it generate code for LangGraph, CrewAI or other frameworks?

Not yet. Framework configuration formats change often, and a generated starter that does not run is worse than none. The JSON export is a plain description of nodes and edges that is easy to map onto any framework.

Is my design saved?

The current design is saved as a draft in your browser and restored when you come back. Nothing is uploaded. Use the share link or the JSON export to keep or send a copy.

Does it work on a phone?

Yes, with touch dragging, though a larger screen is easier for big designs. Every node can also be moved with the keyboard: Tab to it, use the arrow keys, and press Delete to remove it.