In the Agent era, the type system is the hardest feedback
In his Bug Bash 2026 talk, Ron Minsky said one thing: when an Agent writes code, what it needs most is fast, unambiguous feedback. Type systems, expect tests, and formal verification—these old methods are actually more valuable in the Agent era.
Ron Minsky's talk at Bug Bash 2026 was titled “Now more than ever: building reliable software in the age of agents.” At first glance, the title sounds like one of those “how many times do I have to say this before you believe me” talks. But the content was very concrete.
He is from Jane Street. Jane Street is a global trading firm with about 3,500 people, of whom about 2,000 write code every day. Their systems handle roughly 10% of US equities trading volume. The firm has a well-known quirk: its primary language is OCaml, and that wasn't enough, so they added their own extensions to it. Software at a trading firm has to be reliable—the systems are directly connected to the money. Ron mentioned that they started building AI assistant tools several years ago, but for a long time they were a net productivity drain—until the models crossed some threshold last November, “like a sailboat finally being built and then the wind suddenly arrived.”
The core point of the talk: Agents depend on feedback to survive. The type system is the hardest form of feedback. A type system catches errors before the code runs; it doesn't need a test environment, and it doesn't need a human in the loop. For an Agent, a type error is an immediate, free, indisputable review. A commenter in the tweet thread summarized it more crisply than the title: “Types are the only review that an Agent can't sweet-talk its way past. It can't get around a type error; it can get around a flaky test.”—for a test that sometimes passes and sometimes fails, the Agent can always find an excuse to rerun it.
Ron discussed several concrete things.
Make illegal states unrepresentable
make illegal states unrepresentable—an old saying that has circulated in the OCaml community for years. Rather than writing in the docs that “this field cannot be changed while in the refunded state,” it is better to compress the state space to the legal range and let the compiler enforce the global rules. Agents don't need to understand business constraints; the type system blocks the path at the compilation level. This is especially important for Agents: there is much less for them to guess.
Expect tests
Print program output directly and freeze it as a test. Fixture tests in the compiler field have always worked this way. Lauren (@poteto) gave an example in the tweet: React Compiler's tests use a whole set of such fixtures. The benefit is that behavior is readable—during review, you just check whether the output is reasonable, rather than tracing through the implementation logic line by line.
Property testing, fuzzing, deterministic simulation
These methods themselves are not new. But Agents generate large amounts of code, and they may break systems in ways you didn't anticipate. Property testing and fuzzing can automatically find boundary issues. Whole-system deterministic simulation puts the entire system into a controlled environment and runs it there. Relying on humans to watch the code Agents write doesn't scale; you need machines to watch machines.
The cost-benefit of formal verification has changed
seL4 is the classic example: 8,700 lines of C code, 25 person-years, and 23 lines of proof for every line of C. In the past, no matter how you did the math, it didn't pay off. Agents have changed the cost structure. Formal verification used to be expensive because of human labor, and Agents may be able to take on a large amount of the mechanical proof work. This thing may go from an academic luxury to an engineering option.
Ron also raised a point that is easy to overlook: code review has not disappeared in the Agent era; it has become more critical. The focus of review is also shifting. Besides whether the code itself is correct, you also need to see whether it continues the system's design intent.
A tweet reply put it even more bluntly: repos are prompts now. The codebase is the Agent's prompt. The quality of the codebase determines the quality of the Agent's output.
The most counterintuitive part of the whole talk is here: in the AI era, the value of old methods has actually been amplified. Type systems, tests, and review—things that have existed for decades—have become more important infrastructure.
发布时间: 2026-10-27 02:14