Research summary, Sep 2026

Headline

Nobody has built this yet. Mojo hit 1.0 in Aug 2026 but is “for AI” in the sense of running AI workloads. Pel is a Lisp-ish agent-orchestration language. The closest real design is a blog sketch (PACT-Lang) that stayed a sketch. The space is open.

Three findings that shape everything

  • Training-data gravity is real. Dan Luu benchmarked agents building a full zstd decoder and modifying Pandoc across many languages. Obscure languages did badly even when denser. Popularity correlated with correctness. A new language starts with zero corpus: our single biggest risk.
  • Strictness helps models converge. Go-for-agents debate, Dafny vericoding (verification success 68% → 97% in one year), verified Rust via Verus and Aeneas. Agents thrive when the compiler rejects wrong programs loudly and there is one way to do things.
  • The ideal AI language is one humans keep rejecting. Explicit effects, exhaustive result types, contracts, totality. Friction for a human typist, a gift for an agent that iterates in a loop and never tires of annotations.

Smaller things worth stealing

  • Token efficiency matters less than people think at scale (Dan Luu). Don’t contort syntax for it.
  • “Lingering Authority” paper: revocable, time-bounded capabilities for coding agents; separate permission from actual effect. Fits an effect system perfectly.
  • Research trend toward compilers emitting structured diagnostics (category, cause, suggested fix) for agent feedback loops.

Sources

Session 6 addendum (Fable, 13 Sep, from the agent-authoring deep run)

The numbers that bound Mo’s cold start and its diagnostics bet, from prompts-mo-parallel-tracks’s third run.^[raw/research-runs/agent-authoring-research-frontier.pplx.md] A no-resource language gets 9 percent pass@1 against Python’s 79, and on Gleam nine in ten generations did not compile; documentation in context moves the odds 1.2 to 11 times; further pre-training on 28 million tokens of a new language lifts hard tasks to 25 to 30 percent, still far from parity. On the diagnostics bet: the reference measurement is RustAssistant, 74 percent of real compiler errors fixed by a model given structured errors in a repair loop, and prompt shape alone moved ownership errors from 10 to 74 percent; structured test feedback beats structured compiler feedback beats prose. On the counter-position: a Python subset with a new discipline (Quasar) produced about seven times fewer erroneous programs than free Python without losing fluency, and the run’s four tests a new language must pass are on case-against-new-languages. Nothing in the run measures capabilities or deterministic replay for agents head to head; the theoretical case is made and the experiment does not exist, which makes Mo’s programs the first data.