Multi-Agent Collaboration

A single agent can handle straightforward requests. But complex workflows — research pipelines, content production, data analysis — benefit from multiple specialized agents working together. Saltare supports this through multi-agent collaboration.

How Collaboration Works

The Collaborate Tool

Any agent can invoke another agent using the collaborate tool. When this happens:

1. The initiating agent sends a request to the target agent
2. A hidden thread is created for the agent-to-agent conversation
3. The target agent processes the request using its own tools and context
4. Results are passed back to the initiating agent
5. The initiating agent incorporates the results into its response to you

You see the final result in your thread. The agent-to-agent communication happens behind the scenes.

Why Hidden Threads?

Hidden threads keep your channels clean. When a Researcher agent asks an Analyst agent to crunch numbers, you don't need to see every back-and-forth message between them — you just want the final answer.

Designing Agent Squads

A squad is a group of agents with complementary skills that work together. Effective squads have:

Clear Role Separation

Each agent should have a distinct specialty:

Agent	Role	Key Tools
Researcher	Discovery and data gathering	web_search, fetch_url, search_news
Analyst	Data processing and evaluation	query_data_table, natural_language_query
Writer	Content creation and editing	create_document, update_document
PM	Task management and coordination	create_task, update_task, search_tasks

Minimal Overlap

Agents with overlapping capabilities create confusion. If both a Researcher and an Analyst can search the web, it's unclear who should handle a research request. Keep tool assignments focused.

A Coordinator Pattern

For complex workflows, designate one agent as the coordinator:

@PM plan and execute a competitive analysis of Acme Corp.
Use @Researcher for data gathering, @Analyst for scoring,
and @Writer for the final report.

The PM agent orchestrates the workflow, delegating to specialists and assembling the final deliverable.

Collaboration Patterns

Sequential Handoff

One agent completes its work, then passes results to the next:

Researcher → finds data → Analyst → scores data → Writer → creates report

Best for: linear workflows where each step depends on the previous.

Parallel Fan-Out

Multiple agents work on different parts simultaneously:

                → Researcher A (analyzes vendor 1)
Coordinator →   → Researcher B (analyzes vendor 2)  → Synthesizer
                → Researcher C (analyzes vendor 3)

Best for: work that can be divided into independent chunks.

Review Loop

One agent creates content, another reviews it:

Writer → creates draft → Editor → reviews and suggests → Writer → revises

Best for: content quality assurance.

Setting Up Collaboration

Step 1: Create Specialized Agents

Build agents with focused roles. Give each a clear system prompt that defines what it does and doesn't do.

Step 2: Assign Relevant Skills

Limit each agent's tools to what its role requires. A Writer doesn't need database query tools. A Researcher doesn't need task management tools.

Step 3: Test in a Dedicated Channel

Create a test channel and run through a workflow:

@PM I need a brief on our 3 main competitors. 
Coordinate with @Researcher and @Writer to produce 
a comparison document by end of day.

Watch how the agents collaborate, then refine their system prompts based on the results.

Limitations

Context Boundaries

Each agent has its own context window. When Agent A collaborates with Agent B, it sends a summary of what it needs — not its entire conversation history. This means:
- Keep collaboration requests specific and self-contained
- Don't assume Agent B knows what you discussed with Agent A
- Important context should be written to documents (persistent) rather than kept in chat (ephemeral)

Collaboration Depth

Agents can collaborate one level deep — Agent A can ask Agent B for help, but Agent B shouldn't ask Agent C. This prevents runaway chains and keeps workflows predictable.

Iteration Limits

Each agent has a maximum number of tool-call iterations per request (5 by default). Complex collaborations may need to be broken into multiple requests to stay within limits.

Tips

• Start with two agents — A researcher and a writer cover most workflows
• Name agents clearly — Role-based names (Researcher, Analyst) are better than creative names for collaboration
• Document the workflow — Write down which agent does what, so your team can trigger the same pipeline
• Check hidden threads occasionally — Review agent-to-agent conversations to understand how they're collaborating and identify improvements