agent-collaboration

🤝 Agent Collaboration

Philosophy: Multiple specialized agents outperform one overloaded generalist.

When to Use Multi-Agent

Single Agent	Multiple Agents
Focused task	Complex workflow
One context	Multiple contexts needed
Sequential work	Parallel execution benefits
<500 lines changed	Large-scale changes

Collaboration Patterns

1. Hub and Spoke (Orchestrator)

Central agent coordinates specialists:

           ┌── Coder ──┐
           │           │
Orchestrator── Tester ──├── Combine
           │           │
           └── Docs ───┘

Use when: Clear subtask boundaries, need coordination.

2. Pipeline (Sequential Handoff)

Each agent completes then passes to next:

Planner → Coder → Reviewer → Deployer

Use when: Each stage needs different expertise.

3. Debate (Adversarial)

Agents critique each other's work:

Proposer ←→ Critic → Improved Result

Use when: High stakes, quality critical, catching errors matters.

4. Swarm (Parallel Workers)

Multiple agents work on similar tasks:

     ┌─ Worker A (files 1-10) ─┐
Task ┼─ Worker B (files 11-20) ┼─ Merge
     └─ Worker C (files 21-30) ─┘

Use when: Embarrassingly parallel, same task type.

5. Recursive Decomposition (The Thinking Tree)

Based on StructuredAgent (arXiv:2603.05294v1). Every goal is treated as a root node in a dynamic AND/OR thinking tree.

          [ Root Goal ]
          /           \
    [Plan A (OR)]      [Plan B (OR)]
     /        \          /        \
[Task 1] [Task 2]   [Task 1] [Task 2]
 [AND]    [AND]      [AND]    [AND]

Orchestrator Protocol:

• Static vs. Dynamic: Don't just follow a list. If Task 1 of Plan A fails, the orchestrator prunes Plan A and activates Plan B.
• Node Injection: If a new blocker is discovered, the orchestrator injects a new AND child into the tree to resolve the blocker before proceeding.
• State Propagation: Success in a leaf AND node propagates up to its parent OR node, satisfying the parent and concluding that branch of the search.