Tag
A comprehensive walkthrough on building multi-agent teams with Claude Managed Agents, covering role design, model mixing, and parallel execution to scale from one to 20 agents.
The article presents a practical critique of long-running LLM sessions, life-companion agents, and persistent memory systems, raising issues of privacy, cost, intent-loss, and maintainance. It proposes alternative solutions like issue-bound ephemeral session chains and prompt templates.
A controlled study of compound LLM agent design in an adversarial POMDP (CybORG CAGE-2), systematically varying context, reasoning, and hierarchy across five model families. Key findings: programmatic state abstraction yields large returns per token, hierarchy without deliberation tools achieves best absolute performance, and context engineering is more cost-effective than deeper reasoning.
Discusses trade-offs between fixed agent roles and dynamic spawning in multi-agent LLM systems, based on personal experience building a multi-agent setup. Explores when explicit specialists are beneficial versus when they add unnecessary ceremony.