Communication Patterns: Sharing the Context

When multiple agents work together, the biggest risk is "Information Loss." If the Researcher finds 10 facts but the Writer only receives 3 of them, the system has failed.

In CrewAI, communication is handled through Context and Shared Memory.

1. Automated Context Passing

By default, when Task B follows Task 1, Task B receives the Final String Output of Task 1.

The Challenge: What if Task B needs the output of Task 1 AND Task 2? The Solution: Use the context parameter in the Task definition.

# Task 3 now gets the results from BOTH 1 and 2 automatically.
task3 = Task(
    description='Summarize all findings.',
    context=[task1, task2],
    agent=manager
)

---

2. Shared Memory (The Kitchen)

CrewAI has a built-in memory system that allows agents to share "Insights" in real-time.

• Short-Term Memory: Agents share their current task progress.
• Long-Term Memory: Agents can store results in a local database to use in future executions of the crew.

---

3. Communication Verbosity

When debugging multi-agent systems, you need to see the "Conversation."

• verbose=True: Shows the agent's internal monologue.
• step_callback: Allows you to send every action to a dashboard or Slack for monitoring.

---

4. Visualizing the Data Flow

graph TD
    T1[Researcher Output] -->|String| T2[Analyst Input]
    T1 -.->|Memory Store| T3
    T2 -->|String + Context| T3[Writer Input]
    T3 -->|Final Result| User

---

5. The "Human-Friendly" Protocol

Because agents communicate in Natural Language, they sometimes "over-explain" things to each other.

• The Problem: Agent A sends a 5-page report to Agent B. Agent B hits its token limit and crashes.
• The Fix: Use Structured Output (Pydantic). Instead of sending a report, Agent A sends a ReportMetaData object. Agent B then knows exactly where to look for the information without reading 5 pages of fluff.

---

Key Takeaways

• Context Passing is the backbone of multi-agent synergy.
• The context parameter allows for non-linear data gathering.
• Shared Memory helps agents learn across different tasks and sessions.
• Structured Output is the best way to prevent "Token Bloat" during agent-to-agent communication.