Skip to content

The Handoff -- Multi-Agent Shared Memory

Three agents (Researcher, Developer, QA) share a memory store. The Researcher says the API rate limit is 1000 req/min (from documentation). The Developer says it's 500 (from testing, and is frustrated). QA confirms 480 from load tests. Contradiction detection catches all conflicts with agent attribution.

What It Demonstrates

  • Multi-agent shared memory with agent_id tagging
  • Contradiction detection across agents
  • Emotional valence (Developer's frustration as a social signal)
  • Agent attribution in synthesized responses

Key Moment

A new developer asks about the API. Claude synthesizes:

"There is a significant contradiction between the official documentation and real-world testing. The documented limit of 1,000 req/min is wrong. I'd recommend designing your integration to stay comfortably below 480 req/min."

The response includes agent attribution and notes the Developer's frustration as a social signal.

Running

python -m demos.handoff --backend anthropic  # ~2 min, ~$0.30