Canvas Builder Guide
Everything you need to know to build, configure, and publish your first swarm. No coding required.
What is a swarm?
A swarm is a chain of AI agents working together in sequence. Each agent handles one specific task and passes its output to the next agent — like an assembly line for AI.
Researcher
Finds info
Analyst
Processes it
Writer
Creates output
Verifier
Checks quality
Each agent runs a different AI model, has its own instructions, and builds on what came before it. The final agent delivers the completed output to the user.
The six agent types
SwarmSeller has six agent types. Each has a different default role and works best in a specific position in your swarm.
Orchestrator
Start of swarmPlans the task, reads user input, and directs other agents. Always use this as your first node. Think of it as the project manager.
Researcher
Early in swarmSearches the live web using Perplexity. Use this whenever your swarm needs current information, news, job listings, market data, or anything that changes frequently.
Analyst
Middle of swarmProcesses, scores, and evaluates information. Use this to rank options, score opportunities, build frameworks, or make structured decisions from raw data.
Writer
Middle to lateCreates written content — articles, emails, reports, cover letters, social posts. You can use multiple Writers in one swarm for different types of content.
Executor
Middle of swarmTakes action or builds structured outputs — schedules, plans, step-by-step processes, and organised deliverables. Think of it as the doer.
Verifier
End of swarmReviews, fact-checks, and quality-controls the output from all previous agents. Always use this as your last node. It catches errors and ensures the final output is ready to use.
Building on the canvas
The canvas is your workspace. Here is how to use it step by step.
Add nodes from the left panel
Click any agent type in the left sidebar to add it to the canvas. Each click adds one node. Add them in the order you want them to run — top to bottom works best for readability.
Tip: Start with an Orchestrator. End with a Verifier. Fill the middle with whatever your swarm needs.
Connect your nodes
Hover over a node until you see a small circle appear on its edge. Click and drag from that circle to another node to create a connection. Connections show the flow of information between agents.
Tip: You can branch — one node can connect to two nodes running in parallel. Both will receive the same input and run simultaneously.
Configure each node
Click any node to open its configuration panel on the right. Here you can change the name, select the AI model, write the system prompt, set temperature, enable memory, and turn on the human approval gate.
Tip: The system prompt is the most important setting. Write it in plain English — tell the agent exactly what it is, what it receives, and what it should produce.
Delete a node
Click any node to select it (it will highlight), then press the Delete or Backspace key on your keyboard to remove it.
Tip: Connections attached to a deleted node are also removed automatically.
Save your swarm
Click the Save Swarm button in the top right. A modal will appear asking for a name, description, and emoji. Fill these in carefully — they appear on your marketplace listing.
Tip: You can save and come back to edit at any time. Your swarm is saved as a draft until you publish it.
Choosing the right AI model
Each node can use a different AI model. Choose based on what that agent needs to do.
| Model | Best for | Avoid for |
|---|---|---|
| Claude Sonnet | Writing, reasoning, long documents, following complex instructions | Live web data |
| GPT-4o | Data analysis, structured outputs, scoring, fast processing | Very long outputs |
| Perplexity Sonar | Live web search, current news, job listings, market data | Creative writing, reasoning |
| Gemini | Very long documents, multimodal tasks, large context | Nuanced writing |
| Grok | Current events, social media trends, cultural context | Precision tasks |
| Ensemble | When you want the best of multiple models combined | Speed-critical tasks |
Advanced settings
Temperature
Controls how creative or predictable the AI is. Lower temperature (0.1-0.3) = more focused and consistent output, great for analysis and factual tasks. Higher temperature (0.7-1.0) = more creative and varied output, great for writing and brainstorming. Default is 0.7 which works well for most tasks.
Persistent memory
When enabled, the agent remembers context from previous runs with the same user. Use this on the Orchestrator node when your swarm benefits from knowing the user's history — for example a Life Admin swarm that learns your preferences over time. Only enable on one node per swarm, usually the Orchestrator.
Human approval gate
When enabled, the swarm pauses at that node and asks the user to review and approve the output before continuing. Use this on the Verifier node for high-stakes swarms — financial analysis, medical information, or anything where a human check is important. The user sees the output and clicks Approve or Reject.
Writing great system prompts
The system prompt is the most important part of your swarm. It tells each agent who it is, what it receives, and what it should produce. Here is the formula that works.
The formula
You are a senior career coach specialising in tech industry applications.
You will receive a candidate's CV and a job description.
Produce a tailored CV rewrite and a cover letter.
Mirror language from the job description. Quantify achievements. Keep to two pages.
Do not use phrases like I am passionate about or I am a team player.
Common mistakes
- ✗Being too vague — 'Analyse the data' tells the agent nothing. Be specific about what data and what analysis.
- ✗Forgetting to tell the agent what format to use — should it use bullet points, paragraphs, tables, headers?
- ✗Not telling the agent what input to expect — mention that it receives output from the previous agent.
- ✗Making the system prompt too long — keep it focused. One clear job per agent.
Publishing your swarm
Once your swarm is saved and tested, you can publish it to the marketplace. Here is what you need to prepare.
Tips from the SwarmSeller team
Start simple
Your first swarm should be 3-4 nodes maximum. Get it working perfectly before adding complexity. A great 3-node swarm beats a broken 7-node swarm every time.
Test before publishing
Run your swarm at least 5 times with different inputs before publishing. Edge cases will surprise you. Fix them before your users find them.
Specialise your agents
Each agent should do one thing really well. If you find yourself writing a system prompt that does three different things, split it into three separate nodes.
Match models to tasks
Use Perplexity for anything that needs live web data. Use Claude for writing and reasoning. Use GPT-4o for analysis and structured outputs. Mixing models strategically makes swarms dramatically more powerful than using one model for everything.
Write for your user
When writing your example prompt and usage instructions, imagine your least technical user. Make it impossible to get wrong. The more specific your guidance the better your reviews will be.
Price for value not cost
Price based on the value your swarm delivers, not just the token cost. A swarm that saves a user 4 hours of research is worth $20 per run even if it costs you $0.50 to run.
Ready to build?
Open the Canvas Studio and build your first swarm in minutes.