Build1 publisher3 min readPublished
LangChain drops about 4,000 base input tokens from every default Deep Agents turn
The harness lost its hidden system prompt, 43% of its builtin tool descriptions and its todo list middleware. LangChain's own footnote says reward confidence intervals span zero for every model tested, so the evals settle the token saving more firmly than the quality.
The Engineer · Build desk
What happened
- LangChain's Deep Agents v0.7 cuts base input tokens on a default agent turn from about 6,000 to about 2,000, a 65% reduction the company says comes at comparable performance.
- The release removes the hidden base system prompt, shortens builtin tool descriptions by 43%, and stops create_deep_agent from including TodoListMiddleware by default.
- Validation ran v0.7 against the v0.6.12 baseline across autonomous, conversational and long-context benchmarks on gpt-5.6-luna, gemini-3.6-flash, claude-sonnet-4-6 and claude-opus-4-8.
- gpt-5.6-luna used 34% fewer tokens at 15% lower cost with reward up 4%, while claude-sonnet-4-6 got more expensive, a rise LangChain traced to two challenging autonomous tasks.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Every default v0.6.12 turn carried roughly 4,000 tokens of prompt, tool text and middleware that the new evals could not show earning anything. That was billed on trivial calls as well as hard ones.
- constraint Reward confidence intervals span zero for every model, so the suite can demonstrate the token saving and cannot certify that quality held. Anyone whose task mix differs from the three categories carries the regression risk themselves.
- decision Teams on smaller or older models now have to decide whether to re-add planning scaffolding that used to arrive by default, because LangChain says less capable models still need it to avoid dropping steps.
- precedent Two harness authors trimming prompts by 65% and 80% in the same period sets the expectation that prompt text gets re-audited on each model generation, in step with the vendor guides being rewritten.
Base input tokens are what the harness spends before the task starts: the builtin prompt, the tool schemas, the middleware [3]. They are billed on every turn, whether the agent is refactoring a repository or answering a one-line question.
The 65% is a base-token figure. On gpt-5.6-luna, total tokens fell 34% and cost fell 15% [10], so the bill moved at about 0.44 times the rate of the token count [2]. Take the two percentages together and assume the trim was the only change: a 65% cut to base tokens yielding a 34% cut to total tokens puts the old base at roughly half of everything measured in that suite [3].
LangChain's footnote sets the limit on the quality claim. "Reward confidence intervals span zero for every model," the post says, while Luna and Opus show statistically clear token reductions and Luna also a statistically clear cost reduction [12][13]. The firm result is about tokens. For the reward result to carry to your agent, your work has to resemble the three eval categories: end-to-end coding and data analysis, multi-turn conversation with a simulated user, and retrieval and reasoning over long context [7]. And your model wants to be one of the four in the matrix [8].
claude-sonnet-4-6 went the other way, and LangChain says the cost increase came largely from two challenging autonomous tasks it found in LangSmith traces [11].
LangChain kept TodoListMiddleware shipping and says it still helps on long multi-step tasks, on less capable models that drop steps without more scaffolding, and in UI-facing cases where a visible plan matters [18]. Turning it back on is one line, `middleware=[TodoListMiddleware()]` [19]. The planning prompt and `write_todos` tool "did not significantly improve performance" in their evals [6], and those todo evals ran across three categories and three models [17], one model fewer than the harness comparison [8].
The same week, Anthropic reported cutting over 80% of Claude Code's system prompt for models like Opus 5 and Fable 5 with no measurable drop in coding evals [14]. Its two findings match what LangChain saw. The first: "Interfaces beat examples: good tool schemas teach usage better than the once popular few shot examples, which can narrow how the model explores" [15]. The second is that repeating an instruction in both the system prompt and a tool description offers no meaningful reinforcement [16]. That second rule you can check against your own harness by grepping the system prompt for instructions that also appear in tool descriptions. LangChain notes that OpenAI, Anthropic and Google all publish prompting guides and all three have rewritten them as models got more capable [21].
Anyone who built on the old defaults has work to do on upgrade. The hidden base system prompt, with its general guidelines and tool-usage prose, is gone [4]. Instructions written to sit on top of it are now the whole of what the model gets. LangChain says configurability was the top ask from Deep Agents users over the last six months: requests to override FilesystemMiddleware, customize SummarizationMiddleware thresholds and override the base prompt globally. All of them hit the same wall. There was no supported way to change what the default harness stack does [20].
What to watch
- Whether the reward confidence intervals narrow in LangChain's full report or in later eval runs. Narrower intervals are what would turn "no detected change" into equivalence.
- Whether the claude-sonnet-4-6 cost increase reproduces outside the two autonomous tasks LangChain traced it to.
- Whether harness maintainers start shipping per-model prompt profiles as vendor prompting guides are rewritten again.