LangChain strips agent system prompt, cuts tokens 65%

LangChain cut default agent tokens roughly 65% in Deep Agents v0.7, mostly by deleting the base system prompt, per the July 29 release notes.

For anyone running agents at volume, that's real money: a default turn now costs about 2,000 input tokens instead of roughly 6,000. At a thousand turns a day, that's 4 million tokens of headroom back, free.

The same release also moved TodoListMiddleware, the feature that has an agent write down its own plan before acting, from default to opt-in. LangChain's evals found it didn't meaningfully help across most use cases.

LangChain also trimmed tool descriptions 43%, which combined with the prompt deletion to cut tokens, with comparable performance across models.

Two changes matter if you're upgrading: deprecated backend factory support is gone, and v0.7 adds a new delete tool to the default filesystem toolkit.

Hugo Bowne-Anderson made a similar case in an O'Reilly Radar piece on July 22. He credits three moves to Lance Martin: reduce what you pass into the prompt, offload complexity outside it, and isolate token-heavy subtasks into separate agents. His example: OpenClaw handles memory with timestamped Markdown files instead of a vector database.

A separate repo, shareAI-lab/learn-claude-code, makes the same point about skill loading. Ten skills inlined at roughly 2,000 tokens each burn 20,000 tokens most turns don't need. Its fix: list skill names and one-line descriptions in the system prompt, then load the full skill only when the model asks for it.

Turn something off

Pick the most elaborate piece of your own harness, the middleware you're proudest of, and turn it off. Run your evals. Budget an afternoon. If nothing moves, you just made your agent cheaper and faster by deleting code, which is the best trade in software.