Claude Code is getting rid of SUB-AGENTS?
Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join In this video, I'll walk you through Claude's new article, "Building agents: When to fold subagents into your main agent loop", in plain words: why copying your org chart into subagents hurts quality, adds cost and latency, and what progressive disclosure, skills and tools do instead. Then I'll cover how to fold one subagent at a time, how to measure it with a small set of golden examples, and the three reasons the article gives for keeping a subagent separate. In the second half, I'll show you how the same choice looks in PandaCode, with its single chat and its team mode: the architect and workers, per-worker harness, model and effort, team profiles, read-only and writing workers, file scopes and follow-ups. Finally, I'll go through where a team genuinely wins, where one chat is better, the real costs, and my own rule for picking between them. -- Key Takeaways: 🧠 Claude's article argues many teams overuse subagents by copying their org chart into the agent, an idea it ties to Conway's Law. ⚡ Its alternative is one main agent loop that loads skills and tools only when needed, called progressive disclosure. 📊 Anthropic says it cut the Claude Code system prompt for its more advanced models by over 80% in part by disclosing skills and tools this way. ⚠️ The failure modes are subagent sprawl, the Frankenprompt, the telephone game between handoffs and extra latency from every nested loop. 🔀 Fold back one subagent at a time: prompt into a skill, API calls into tools, and measure before touching the next one. 🔍 Even 10 golden examples is enough to start, and run each eval more than once, because one folding break showed up in only one run out of two. 🎯 Keep a subagent separate only if you can measure that its context hurts the main agent, if work runs well in parallel in fresh contexts, or if one agent keeps struggling with tools or conflicting instructions. 👥 In PandaCode, the architect plans and dispatches workers that each run in a fresh context, with their own harness, model and effort, and saved team profiles. 💸 A team is not free speed: every worker is another model run, summaries lose some detail, and workers share one checkout, so file scopes matter. ✅ My rule: default to one chat, earn the team with a reason you can name, then start simple, measure, change one thing at a time and keep what holds up.
