Templates

When you describe a task to the AgenTopology skill, it selects the best architecture pattern for your use case. These are the patterns the skill draws from.

Available patterns

Pipeline

Best for tasks with a clear sequence: research, then write, then review. Each agent runs in order, and conditional edges can create retry loops.

topology example : [pipeline] {
  agent step-one   { phase: 1 }
  agent step-two   { phase: 2 }
  agent step-three { phase: 3 }

  flow {
    intake -> step-one -> step-two -> step-three -> done
  }
}

The skill picks this when your description mentions steps that happen "in order," "sequentially," or "one after another."

Supervisor

Best for tasks where a central coordinator needs to route work to different specialists based on the input. The supervisor decides which agent handles each request.

topology example : [supervisor] {
  orchestrator {
    model: opus
    handles: [intake, done]
  }

  agent specialist-a { role: analyst }
  agent specialist-b { role: writer }
  agent specialist-c { role: reviewer }

  flow {
    intake -> orchestrator
    orchestrator -> specialist-a [when type == analysis]
    orchestrator -> specialist-b [when type == writing]
    orchestrator -> specialist-c [when type == review]
  }
}

The skill picks this when your description mentions "routing," "delegation," or tasks that need different handling based on input type.

Fan-out

Best for tasks where the same input needs to be processed by multiple agents simultaneously. Results are aggregated after all agents complete.

topology example : [fan-out] {
  agent analyzer-a { role: security-review }
  agent analyzer-b { role: performance-review }
  agent analyzer-c { role: style-review }
  agent aggregator  { role: summarizer }

  flow {
    intake -> analyzer-a
    intake -> analyzer-b
    intake -> analyzer-c
    analyzer-a -> aggregator
    analyzer-b -> aggregator
    analyzer-c -> aggregator
    aggregator -> done
  }
}

The skill picks this when your description mentions "parallel," "simultaneously," "multiple reviewers," or "independent analyses."

Blackboard

Best for tasks where agents need shared context. A central knowledge store (the blackboard) is read and written by all agents. Agents can run in any order and build on each other's findings.

topology example : [blackboard] {
  memory shared {
    path: "workspace/blackboard.md"
    access: read-write
  }

  agent researcher { reads: ["workspace/blackboard.md"] writes: ["workspace/blackboard.md"] }
  agent analyst    { reads: ["workspace/blackboard.md"] writes: ["workspace/blackboard.md"] }
  agent synthesizer { reads: ["workspace/blackboard.md"] writes: ["workspace/report.md"] }
}

The skill picks this when your description mentions "shared knowledge," "building on findings," or "collaborative analysis."

Pipeline + Blackboard

The most common pattern for production systems. Agents run in sequential phases but share a common workspace. Each agent reads previous agents' contributions and adds its own.

topology example : [pipeline, blackboard] {
  memory shared {
    path: "workspace/context.md"
    access: read-write
  }

  agent researcher {
    phase: 1
    writes: ["workspace/context.md", "workspace/research.md"]
  }

  agent architect {
    phase: 2
    reads: ["workspace/context.md", "workspace/research.md"]
    writes: ["workspace/context.md", "workspace/design.md"]
  }

  agent builder {
    phase: 3
    reads: ["workspace/context.md", "workspace/design.md"]
    writes: ["workspace/context.md", "workspace/output/"]
  }

  flow {
    intake -> researcher -> architect -> builder -> done
  }
}

The skill picks this when your description involves multiple steps that need to share context, which covers most real-world agent teams.

How the skill decides

The skill analyzes your description for signals:

• Sequential keywords (then, after, next) point to pipeline
• Routing keywords (depending on, based on, route) point to supervisor
• Parallel keywords (at the same time, independently, in parallel) point to fan-out
• Shared state keywords (shared context, build on, collaborative) point to blackboard
• Complex workflows with both sequence and shared state get pipeline + blackboard

You can always override the skill's choice by specifying the pattern you want.

What's next

• Learn the language reference for all available blocks and fields
• See real-world examples of these patterns in action
• Read about validation rules that apply to each pattern