A reasoning sandwich is a heuristic, named in LangChain’s
Improving Deep Agents with harness engineering, for deciding how much reasoning compute
an agent spends at each stage of a task: more at the start, for planning, and at the end, for
verification, and less for the implementation between them. LangChain’s version, for a model offering
low, medium, high and xhigh reasoning modes, was xhigh for planning, high for implementation
and xhigh for verification.
Usage
The term comes from harness design for long-running coding agents, where reasoning models can run for
hours and a harness has to decide how much compute each subtask gets. LangChain’s reasoning was that
more reasoning helps an agent fully understand a difficult problem and build a good plan, and helps
late-stage verification catch mistakes before a solution is submitted — but that it can also burn over
twice the tokens and time. Under Terminal Bench’s time limits, running the whole task at xhigh scored
53.9% because agents timed out, against 63.6% at high; the team reports no large differences between
different splits of the reasoning budget in trial runs, and kept the sandwich as its baseline.
When It Applies
It assumes a model whose reasoning effort can be set per call (see Reasoning Effort) and a task with a binding time or cost budget, which is what made spending maximum reasoning everywhere counterproductive in LangChain’s setting. It is one team’s heuristic from its own benchmark runs, not an established practice. The same post names adaptive reasoning — as in Claude and Gemini models, where the model itself decides how much to reason — as the natural alternative, and suggests that in a multi-model harness the same balancing could instead mean planning with a large model and handing implementation off to a smaller one.