Seven bets on where agents go next — notes on the LangChain Interrupt closing keynote

These are my notes on a talk by someone else — Harrison Chase’s closing keynote at LangChain’s Interrupt conference. The framing he chose was to stand a year in the future, at Interrupt 2027, and ask what the industry would be arguing about by then. He was explicit that he expected to be wrong in places. Summarised here in my own words; go to the original for what he actually said.

Agents split into two kinds. Long-horizon agents that run for minutes or hours — code execution, planning, sub-agents — and latency-sensitive ones where response time decides everything, mostly customer-facing: support, sales, voice. They share a stack underneath, but how much of that stack is genuinely common is an open question.

Voice becomes a real topic. Today’s pipeline is a sandwich: speech to text, a text-space agent, text back to speech. Native speech-to-speech models are arriving but are not yet steerable enough for applications that care about control. Expect that to change, and expect the pipeline-versus-native argument to run for a while.

Every agent needs a sandbox. Because writing code turns out to be a general-purpose tool — not just for software, but for data analysis, browsing, research. His framing stuck with me: giving an agent a sandbox is like giving a marketing team its own software engineer, and asking what that engineer would build for them.

Open models grow. Three reasons: base models are now close to the frontier without any targeted post-training; token cost is becoming a real constraint, especially for coding agents; and open weights are what let you post-train on your own domain.

Agent identity is unsettled. An agent either acts with your credentials — so two colleagues asking the same agent get different answers, because they can see different things — or it holds a fixed service account and answers everyone identically. Both will exist. What matters is being precise about which one you are running, and telling users which one they are talking to.

Continual learning, in three layers. An agentic system is the model, the harness (the code between model and environment) and the context you feed the harness. All three can be improved. The part I keep coming back to: when you improve the harness or the context, your evals play the role gradients play in classical machine learning. That makes evals and traces the assets, not the paperwork.

Everyone builds agents. Domain experts already give the best feedback about how an agent should behave. The bet is that they stop handing that feedback to an engineering team and start building the agents themselves.

The keynote closed with a launch along those lines — LangSmith Fleet, for building agents in natural language — demoed live on a go-to-market agent, with a human-in-the-loop step before anything is sent to a real customer.

Why this is in my library: the three-layer split is the most useful thing I took away. It gives you somewhere specific to put an improvement, instead of the vague instinct to “make the agent better”. I use the same split when I look at my own automation work.