Shared vs. Private Context: Managing the Common Ground

When multiple agents cooperate, they need a "Common Operating Picture." If one agent learns a crucial fact (e.g., "The server is down"), the other agents shouldn't have to re-discover it.

However, if you share everything with every agent (The "Global Shared Context"), you create a Context Avalanche. Every turn for every agent becomes massively expensive as they read about facts they don't need to do their specific sub-job.

In this lesson, we learn Selective Knowledge Sharing. We’ll differentiate between Private Context (Agent-specific) and Shared Context (Global), creating a high-efficiency information ecosystem.

---

1. The Strategy of "Need-to-Know"

Information in a multi-agent system should be treated like a security clearance.

• Private Context: The agent's own internal reasoning, intermediate tool outputs, and draft work. (90% of data).
• Shared Context: Final decisions, discovered system-wide facts, and changes to the user's core request. (10% of data).

The Goal: Keep the Shared Context small so that it can be injected into all agents without breaking the bank.

---

2. The Shared State Hub (Broadcast Pattern)

Instead of passing data directly, use a Broadcast Hub.

1. Agent A discovers an important fact.
2. Agent A outputs: SHARE: DB_TIMEOUT=5sec.
3. Supervisor adds this to the Global State.
4. Agent B is invoked. Its prompt includes: System Update: [DB_TIMEOUT=5sec].

graph TD
    A1[Researcher] -- Private History --> A1
    A2[Coder] -- Private History --> A2
    
    A1 -->|Signal| HUB[Shared State Hub]
    A2 -->|Signal| HUB
    
    HUB -->|Brief Prefix| A1
    HUB -->|Brief Prefix| A2

---

3. Implementation: Selective State Merging (LangGraph)

In LangGraph, you can define State fields that are "Local" or "Global."

Python Code: Merging logic

def researcher_node(state):
    # This node has its OWN private local state for searching
    local_history = state.get("researcher_private_log", [])
    
    # ... performs search ...
    
    # We return the new private log BUT also a 'Global Summary'
    return {
        "researcher_private_log": local_history + ["Found SKU-123"],
        "global_knowledge": state["global_knowledge"] + ["SKU-123 is BLUE"]
    }

---

4. Pruning the "Shared" Context

Even Shared Context can grow too large. You must periodically De-duplicate it.

• Before: "The meeting is at 5pm", "The room is 302", "Actually, meeting is 6pm."
• After (Consolidated): "Meeting: 6pm, Room 302."

By performing a "Consolidation Pass" on your shared hub every few turns, you ensure that every agent in the fleet stays efficient.

---

5. Token ROI: Selective vs. Global

In a 5-agent system with 10 turns each:

• Global approach: Every turn includes ALL previous 50 msgs. Total: 250,000 tokens.
• Selective approach: Every turn includes 5 local msgs + 3 global facts. Total: 25,000 tokens.
• Savings: 90%.

---

6. Summary and Key Takeaways

1. Private by Default: Don't share raw tool logs or intermediate thoughts.
2. Signal-Only Sharing: Only broadcast the "Final Fact" to the fleet.
3. Periodic Consolidation: Prune and de-duplicate the shared state regularly.
4. Agent Focus: Smaller prompts lead to better focus and fewer "Hallucinated Logic" errors in specialists.

In the next lesson, Cost Attribution in Multi-Agent Graphs, we look at چگونه to track which agent is "The Expensive One."

---

Exercise: The Knowledge Filter

1. Imagine a team of 3 agents: a Designer, a Coder, and a Writer.
2. The Designer finds that the user likes the "Dark Theme."
3. The Coder finds a "Syntax Error in line 45."
4. Which findings should be Global?
   • Hint: Does the Writer need to know about the syntax error?
   • Does the Coder need to know about the Dark Theme?
   • Answer: Dark Theme is Global; Syntax Error is Private to the Coder.
5. Calculate the tokens saved if the Writer doesn't have to read about line 45.

---

Congratulations on completing Module 12 Lesson 4! You are now a knowledge architect.