Harrison Chase Keeps Deleting His Own Product

The product that erases itself

Harrison Chase, the founder of LangChain, told 20VC that building agents is hard because large language models are non-deterministic black boxes, and those black boxes create cascading errors when chained into multi-step systems. From that starting point he drew an unusual business conclusion: as the base models improve, LangChain's job is to erase big parts of its own product. "The main way our business changes as the model gets better is we basically have to erase big parts of our product." m:…0896

His template is the shift from JSON-parsing tool calls to native tool calling: developers built a layer of hacks to coax structured output from models, then the models learned to do it natively and the hacks went in the bin. Things a team had to build six months ago, he said, no longer need building — which is why, in his view, the durable asset for a startup is not any given feature but the iteration loop itself. m:…0896

Routing belongs in the harness

Chase's sharper argument was about where agent decisions get made. Model routing, he said, works better at the harness level than the gateway level because the harness sees the task state. "I think we probably believe it belongs in the harness." Routing is not a commodity: building it well is still hard even today, because the models keep changing underneath it. Most of what passes for routing on platforms like OpenRouter, he added, is really just gateway access — getting to the models, not intelligently choosing between them. m:…0896

His mental model of an "open and controllable agent" has three parts: the model, the harness, and the context. The model sets the ceiling, but the harness and the context are where an agent's domain value lives — which is why he thinks value will accrue to whoever writes the best verifiers and captures the deepest domain knowledge. It is a direct challenge to the reflex of picking a model and shipping a demo: the demo is disposable; the harness is the product. Whether the loop itself should be a post-trained or a frontier model, he said, remains an open question. m:…0896

Memory is still the graveyard

The one layer Chase has given up on, for now, is memory. "I mean, I think this memory layer for agents is still massively unsolved." LangChain launched an alpha of a commercial memory solution and then killed it, concluding it had tried to be too general. The hard part, in his telling, is not storing observations — it is distilling them into information worth keeping. The company's current approach is proprietary plumbing: an internal database called SmithDB for agent traces, and an iteration loop where evals and traces drive every change to context, harness, and model. As Chase put it: "The thing that lets you iterate on each of those is evals and traces". But he also warned that building evals is a very hard job, and academic benchmarks are mostly "benchmark maxed" and less useful than domain-specific ones. m:…0896

The cost paradox — and the banks' plumbing

Then came the bill. Chase said coding agent costs have skyrocketed for many organizations; his own teams spend roughly 30 percent of their engineering salary budget on coding tools: "We spend probably roughly 30% of our engineering salary on coding tools." And the cost of AI intelligence and tokens has not fallen significantly over time, contrary to the usual expectation. Multi-agent systems and agent-to-agent traffic are pushing token consumption up by orders of magnitude relative to human API use. His forecast was blunt: organizations get smaller while coding-agent spend catches up to — then exceeds — human salary costs. That is why LangChain built an in-house cloud coding agent, Open Suite, to control its own spend. m:…0896

The All-In hosts supplied the paradox that explains why the bill keeps rising: 95 percent of code is now written by coding assistants, and demand for human coders is still increasing. "Prices, software prices will definitely go down. But the amount of software is going way, way up. It's the Jevons paradox of software." With assistants making bespoke, single-use enterprise software economically viable for the first time. Startups, they argued, are responding by building moats from proprietary data and vertical feedback loops and by pricing on outcomes rather than tokens. Coinbase supplied their counter-example: a fine-tuned 9-billion-parameter open-weight model cut fraudulent transactions by about 30 percent while running faster and cheaper than Anthropic's Opus 4.5. m:…1487

And while the industry argues about models, the banks are building the pipes. Charles Schwab's Rick Worcester told Bloomberg Surveillance that the firm is blocking unauthorized AI agents from touching client accounts — tens to twenties of attempts per week, by the company's own report — while standing up an official pathway: an MCP server for third-party agents and an in-house assistant called Charlie. Schwab says 77 percent of U.S. clients now use AI somewhere in their financial lives, and client engagement in equities and fixed income is at record levels. It is Chase's thesis in institutional form: the model is interchangeable; the harness — and who controls it — is where the business lives. m:…9711