LangChain released Deep Agents v0.7 on July 29, and its core claim is that a default turn of the agent now spends 65 percent fewer base input tokens, roughly 6,000 down to about 2,000, while performance held steady. Base input tokens are the fixed cost baked into every call: the system prompt, tool schemas, and middleware text that get sent before the model sees a single word of the actual request. Because an agent loop stacks up dozens of turns per task, that fixed cost repeats dozens of times over, and it multiplies faster than the useful work does.
LangChain is not citing a third party here. The 65 percent figure is the company’s own internal measurement, taken on the default configuration of the harness on a single agent turn, not an independent audit. That distinction matters for how much weight the number should carry outside LangChain’s own tests.
Three edits produced the reduction. Engineers deleted the default system prompt that Deep Agents shipped internally, which had carried its own set of instructions and tips on how to call each tool correctly. They cut the wording inside builtin tool descriptions by 43 percent. And they flipped the planning layer, TodoListMiddleware, from on by default to opt in, after internal testing found its planning prompt and write_todos tool rarely moved results enough to justify the extra tokens.
LangChain checked whether stripping that scaffolding cost anything, using a new evaluation suite split into three buckets: end-to-end work such as coding and data analysis, back-and-forth exchanges with a stand-in user, and tasks that require pulling facts out of long context windows. The team compared v0.7 against the previous release, v0.6.12, running each bucket on four separate models: OpenAI’s gpt-5.6-luna, Google’s gemini-3.6-flash, and two Anthropic models, claude-sonnet-4-6 and claude-opus-4-8.
Scores stayed roughly flat across the board, and token and cost counts fell in most cases. gpt-5.6-luna posted the biggest swing: 34 percent fewer tokens, 15 percent lower cost, and a reward score that climbed 4 percent. claude-sonnet-4-6 moved the other way on cost, which LangChain attributes to two unusually difficult autonomous-task runs in the test set skewing that model’s total. On the scoring side, none of the four models cleared a bar high enough to call the change a statistically confirmed win; only the token drops for Luna and Opus, and the cost drop for Luna, held up as distinct from noise.
The release lands days after Anthropic published its own refreshed guide on context engineering, in which the company said it had trimmed more than four-fifths of the instructions Claude Code hands to newer models such as Opus 5 and Fable 5, without a measurable hit to coding evals. LangChain says two of its own conclusions line up with what Anthropic reported: a well-built tool schema teaches a model how to use a tool better than stuffing the prompt with few-shot examples does, and stating the same rule twice, once up top and again inside a tool’s own text, does not make the model any more likely to follow it than saying it once.
That is two separate engineering teams reaching the same diagnosis inside the same month. Agent spending tracks the size of the repeated scaffolding wrapped around every call more closely than it tracks the sticker price of the model itself. A team that wants to know if this applies to its own stack can run one check: take a typical call, isolate what share of its input tokens comes from the system prompt, tool definitions, and injected middleware rather than the actual task content, then multiply that share by how many turns an average task takes. A high number means most of the budget goes to scaffolding that gets resent, not to new reasoning.
Version 0.7 also unlocks settings that were previously hardcoded. Developers can swap out individual middleware pieces, including the point at which SummarizationMiddleware kicks in (previously fixed at 85 percent of the context window), and can set custom base prompts without hidden defaults interfering underneath. Several filesystem tools got smaller upgrades. A second call to write_file on the same path now replaces its contents cleanly, where the tool used to fail outright. read_file, when reading a file in pages, now tells the caller how many lines are left and where the next page starts. And grep stops short at 1,000 matches, streaming what it finds instead of stalling when it hits a large directory tree.
Anyone running Deep Agents in production should measure their own token and cost totals after upgrading rather than assume the 65 percent figure transfers directly, since LangChain’s own tests showed the outcome varied by model. Teams building agent harnesses elsewhere should run the same overhead audit before reaching for a cheaper model as the fix.
LangChain detailed the changes and eval results in a company blog post published July 29, 2026.
