Agent Topologies and Diagrams

Full content

## Why topology matters

As soon as you have more than one agent, the problem becomes architecture, not prompt writing.

## Topology 1: Planner → Specialist → Verifier → Human

```text
User request
   ↓
Planner / Chief-of-staff
   ↓
Specialist
   ↓
Verifier
   ↓
Human approval (if needed)
```

Best for:
- important work
- external communication
- content or decisions with risk

## Topology 2: Fan-out / Fan-in

```text
Main task
   ↓
3 parallel analysts
   ├─ Region analyst
   ├─ Product analyst
   └─ Channel analyst
   ↓
Reducer / Synthesizer
```

Best for:
- market analysis
- research across multiple dimensions
- parallel decomposition

## Topology 3: Builder + Guard rail

```text
Task
   ↓
Builder
   ↓
Fact-check / QA / Schema validation
   ↓
Final output
```

Best for:
- code generation
- content generation
- structured outputs

## Topology 4: Memory-aware personal system

```text
Capture
   ↓
Processing
   ↓
Memory store
   ↓
Orchestrator + specialists
   ↓
Summaries / actions / reminders
```

Best for:
- personal assistant systems
- family support agents
- cross-session continuity
- CoachOS-like systems

## How to choose a topology

Ask:
- how risky is the task?
- how parallel is the work?
- does memory continuity matter?
- does a human need to stay in the loop?

## Design rule

Use the simplest topology that can do the job. Complexity is not a virtue.

## References
- AgenTopology example: fan-out market analysis
- Anthropic subagent patterns
- Open orchestration ecosystems