Pentagon’s core strength is running multiple Claude Code agents simultaneously. Each agent works in its own isolated environment, letting you parallelize development across features, projects, and ideas.
Why Multiple Agents
Running several agents at once gives you three advantages:
- Parallelize work — assign different features to different agents and build them concurrently
- Isolate changes — each agent operates in its own git worktree, so changes never collide
- Separate context — each agent maintains its own conversation history and focus area
A single developer working with 3-5 agents can accomplish what would normally require a small team.
Git Worktree Isolation
Pentagon uses git worktrees to give each agent its own working copy of your repository.
How It Works
| Agent | Branch Behavior |
|---|
| First agent in a repo | Uses your existing branch as-is |
| Subsequent agents | Auto-creates a new worktree with a unique branch |
GitWorktree helper automatically:
- Creates a new branch (named after the agent)
- Sets up a worktree in a subdirectory
- Points the agent’s working directory at the worktree
Each agent sees the full repo history but writes to its own branch. Merge when ready.
Worktree isolation means you can have one agent refactoring auth while another builds a new API endpoint — no merge conflicts until you choose to combine the work.
Merging Changes
After an agent finishes its work:
- Review the changes in the agent’s branch
- Merge into your main branch using your normal git workflow
- Terminate the agent when done — Pentagon cleans up the worktree
Workflow Patterns
Parallel Features
Assign each agent a separate feature or task. Example setup:
| Agent | Task | Branch |
|---|
| api-auth | Build authentication endpoints | feat/auth |
| ui-dashboard | Create dashboard components | feat/dashboard |
| test-suite | Write integration tests | feat/tests |
Multi-Project
Point agents at different project directories entirely. You might have:
- One agent working on your backend API
- Another on your frontend app
- A third updating your infrastructure config
Pentagon manages them all from one canvas.
A/B Exploration
Spawn two agents on the same problem with different approaches. Compare their solutions and pick the better one — or combine ideas from both.
- Clone an agent using the + indicator on adjacent cells
- Give each agent different instructions
- Review both approaches in their respective terminals
Specialized Teams
Create agents with specific roles using their Soul (SOUL.md):
| Agent Role | Soul Instructions |
|---|
| Architect | ”Design APIs and data models. Do not write implementation code.” |
| Implementer | ”Write production code based on the architecture docs.” |
| Reviewer | ”Review code changes. Focus on security, performance, and correctness.” |
Organizing Agents
Pods (Spatial Grouping)
Pods are colored rectangles on the canvas that group agents by proximity:
- Select the Pod tool from the bottom-left toolbar
- Click and drag to draw a rectangle around a group of agent desks
- Name the pod (e.g., “Backend”, “Frontend”, “Infrastructure”)
Agents inside a pod’s bounds are automatically associated with it. Move an agent’s desk into or out of a pod to change membership.
Teams (Logical Grouping)
Teams group agents by role or purpose, independent of canvas position:
- Click New Team in the toolbar or sidebar
- Choose an emoji identifier and name (e.g., “Backend API”)
- Assign agents to the team
Teams appear in the sidebar for quick navigation. Use them when agents working on related tasks are spread across different areas of the canvas.
Scaling Tips
Comfortable Range
- 1-3 agents — easy to monitor, good for focused work
- 5-10 agents — productive parallel development, the sweet spot for most workflows
- 10+ agents — powerful but requires more attention to organization
Cost Awareness
Each active agent makes API calls to Anthropic. More agents means higher API usage. To manage costs:
- Use Haiku for simpler tasks (faster and cheaper)
- Use Opus for complex reasoning tasks
- Terminate idle agents you no longer need
- Set agents to Dormant when you want to preserve their session without active API usage
Pentagon is a native macOS app, so the canvas stays responsive even with many agents. The limiting factor is typically API throughput, not local resources.
Next Steps
