How Much Context Should Your Agents Share?

Why this matters

Multi-agent context sits on a spectrum from fully shared (every agent reads/writes one blackboard) to fully isolated (agents communicate only through explicit messages). Shared context maximizes coordination but blows up budget and couples agents; isolated context keeps each agent lean and decoupled but risks duplicated work and lost coordination. Most real systems are in between, and choosing where on the spectrum each piece of state lives — global, group, or private — is the core design decision. Defaulting to 'share everything' is the most common and most expensive mistake.

Demo

The demo implements a tiered context store: global (all agents), group (a subteam), and private (one agent). Each agent assembles its window from the tiers it's entitled to, so coordination happens only where needed.

Try it yourself

1. Inspect each agent's view and confirm private state never appears in another agent's window.
2. Promote a private finding to group scope when two agents need it, and observe the budget impact.
3. Find a value sitting in global scope that only one agent uses — demote it to private to save every other agent's budget.
4. Add a write-permission check so an agent can only write to tiers it owns.

References

• Anthropic — Building effective agents
• Blackboard system (architecture pattern)

Chat about this lesson