DON'T BUILD MULTI-AGENTS

Summary from the TLDR Web Dev newsletter

Multi-agent frameworks for LLM agents aren't great, as they have disappointing performance and fragility. Instead, a "Context Engineering" approach is better, where it's needed to share full agent traces to maintain context and avoid conflicting decisions. For the best agentic experience, start with a single-threaded linear agent for simplicity and later incorporate a new LLM to compress history for longer contexts.