Praxist beat a Claude Code baseline on MLE-bench at one twelfth the model spend, and the mechanism is copyable

$3,054 against $38,370. Same benchmark, better score.

Praxist (arXiv 2608.25955, submitted August 26) replaces per-attempt agent memory with a typed evidence graph of findings, plus lane-structured frontiers and agendas, so later attempts inherit validated mechanisms rather than rediscovering them. On the standardized 75-task MLE-bench suite, the official grader gave Praxist 60 medals for an 80.0% rate, 49 of them gold. A Claude Code baseline running Opus 4.8 got 55 medals at 73.3%, 34 gold (arXiv 2608.25955).

The medal gap is real but modest. The spend gap is twelve to one.

What makes this useful rather than just another benchmark post is that the mechanism is stated plainly enough to steal. Most agent memory is a transcript: everything the agent did, compressed, handed to the next attempt. Praxist stores typed findings with structure, so attempt seven knows that attempt three already validated a preprocessing approach and doesn't burn tokens re-deriving it. The lane structure keeps parallel explorations from collapsing into each other.

This isn't isolated. JIT-Agent (arXiv 2608.25593) trains a model to generate agent harnesses on demand under a fixed four-module protocol covering memory, planning, action protocol and tool orchestration. With it attached, DeepSeek-V4-Flash passes GPT-5.6 on DeepSearchQA by 9.1 points and GLM-5.2 gains up to 20.2. The authors report the generated harnesses are competitive with mature runtimes including OpenCode and Claude Code (arXiv 2608.25593). And OpsHarness (arXiv 2608.25661) opens with a finding SREs should sit with: a general-purpose agent like Codex or Claude Code now often outperforms a purpose-built root-cause-analysis agent, so the remaining gap lives in the harness. Their control plane contrasts successful and failed diagnoses, converts the difference into atomic proposals, and admits updates only through dual-gate verification. It reaches 59.0% top-1 accuracy, 63.4% above the bare general agent (arXiv 2608.25661).

Three papers in one week, all arguing the model is no longer where the returns are.

The caveat I'd attach: MLE-bench is machine learning engineering, which is unusually well-suited to structured evidence accumulation because experiments produce clean numeric feedback. I don't know whether a typed evidence graph helps as much on ambiguous product work where "validated" is a judgment call. Nobody has shown that yet.

What I'd do this week is smaller than rebuilding your loop. Look at what your agent's memory carries between attempts. If it's a compressed transcript, you're paying to re-derive things you already know. Typing even a few findings, "this approach failed for this reason," "this preprocessing is validated," is the cheap version of what Praxist is doing at $3,054.