Anthropic changes its advice: do not use specialized subagents
Anthropic publishes a post in its resources section. Its advice: fold your subagents into the loop of the main agent. One year ago, the usual practice was different. The main agent sent work to subagents with a role: product, backend, frontend or QA. Each subagent did its part and gave a handoff to the next one.
That architecture made sense because of the limits of the models. The context was 100,000 or 200,000 tokens, and it filled up fast. Agents did not work for more than 2 or 3 minutes at a time, and you had to type "continue". Subagents automated that "continue". For Pablo, they also copied the org chart of the company: Conway's law applied to agents.
Since Opus 4.5, the main agent has more context and works for longer in a single loop. Now the problem with subagents is the context. Each subagent starts with a clean context and repeats the code search that the main agent already did. According to the post, Anthropic removes its internal subagents and uses half the tokens at twice the speed.
The advice: if you already have subagents, remove them one at a time and run evals for each change. Move the instructions from their system prompt to documentation or to skills, with progressive disclosure. General-purpose subagents still make sense to find context or to run work in parallel, and Claude Code, Codex and OpenCode already launch them by default. With Pi, you must ask for them explicitly. Watch in the video.

All the news we covered