Interesting Finds — 2026-08-27

A short roundup of items flagged today. Each is a separate find, collected here so the site keeps a single record while each note can circulate individually.

---

1. JIT-Agent — just-in-time harness synthesis for any LLM

National University of Singapore (Zhang et al., 26 Aug 2026) proposes a trainable harness-intelligence model that synthesizes task-adaptive agent harnesses on the fly for off-the-shelf LLMs, rather than tuning the model itself.

Their claim: agent performance is dominated by harness design — memory, planning, action protocol, tool and skill orchestration — and that design remains manual and unscalable. JIT-Agent is formalized as a composable artifact governed by a fixed four-module protocol, trained to customize a harness for a task, repair it for stable execution, and self-evolve by distilling performance signals from an expanding archive of prior harness configurations.

Reported deltas: with JIT-Agent as helper, DeepSeek-V4-Flash surpasses GPT-5.6 on DeepSearchQA (+9.1) and OdysseyBench (+4.3), while GLM-5.2 gains up to +20.2 points. Across controlled evaluations, generated harnesses are described as competitive with mature runtimes such as OpenCode and Claude Code, and as consistently improving multi-scale families of DeepSeek V4, Mimo-V2.5, and Qwen3.6. To be reproduced outside the authors’ harness. If it holds, it is a compounding axis orthogonal to model scaling: harness quality that improves with use rather than only with larger weights.

Source: arXiv:2608.25593 — JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution · paper page huggingface.co/papers/2608.25593 · project bingreeky.github.io/JIT-site · code github.com/bingreeky/JIT · model JIT-Agent/jit-27b

---

2. Michigan Robotics — the department open-sources its program

The University of Michigan Robotics Department (@michiganrobotics, 3.6k followers, 15 public repos) publishes its core curriculum on GitHub.

Pinned courses tell the shape:

• ROB 101 — Computational Linear Algebra (michiganrobotics/rob101, 1.3k stars, GPL-3.0 for code / CC BY-NC 4.0 for text and video) — first-year, first-semester pilot from Fall 2020. Linear systems and matrix methods taught through robotics and intelligent systems, with full lecture, recitation, and lab videos on YouTube, notes, and three projects included.
• ROB 501 — Mathematics for Robotics (rob501, 2.9k stars) and ROB 311 — How to Build Robots and Make Them Move (rob311, 1.1k stars), plus ROB 201, ROB 204, and textbooks mirrored to their own repos.

The useful signal is not a single paper but an entire department treating curriculum as open infrastructure. You can audit the sequence from computational linear algebra through advanced mathematics and build, with code, videos, and homework in the open. The rest of the org (robotics-website, mbot, etc.) shows how the teaching stack is maintained in public.

Source: github.com/michiganrobotics · ROB 101 at michiganrobotics/rob101

---

3. CIRU DualView — exact phase-specialized weight representations

CIRU has published the DualView technical whitepaper as a reproducible repository rather than a PDF-only drop.

From the README: This repository contains the paper, evidence register, supporting research notes, chart data, figures, bibliography, and reproducible build sources for CIRU DualView. Structure is whitepaper/ (manuscript, HTML, EVIDENCE.md, data, figures, scripts, sources), evidence/, and output/pdf/. Read paths point to the Technical whitepaper (PDF), Evidence and provenance register, manuscript source and standalone HTML, with rebuild documented in whitepaper/README.md.

Implementation is linked separately: ciru-ai/ROCmFPX and model dualview-ornith-35b-v1. The title — exact phase-specialized weight representations — implies weights split by inference phase rather than a single quantized artifact. The register-first, rebuildable form is the point: claims are tied to versioned data and figures, not to a slide.

Source: github.com/ciru-ai/dualview-whitepaper

---

4. ESP-DashboardPlus — a real-time, on-device dashboard for ESP32

aaronbeckmann/ESP-DashboardPlus (337 stars, MIT) is a real-time, on-device web dashboard library for ESP32, ESP32-S3, and ESP32-C3. It serves its own UI over WebSocket and fits the entire frontend — dashboard, console, and OTA — into a single gzip-compressed HTML file stored in PROGMEM (~22KB).

Feature set is deliberately complete for an embedded device: stats, charts with multiple lines and legends, gauges, toggles, sliders, buttons, time and location pickers, collapsible grouped cards with weight-ordered layout and grid spanning, light and dark themes, and desktop and mobile support. A dedicated firmware-update tab carries device info, and a full-page console tab adds filtering, export, and command input. Theming supports custom titles and brand overrides. Install is lib_deps = ESP-DashboardPlus in PlatformIO, or manual drop-in.

It is the kind of project that makes a lab ESP useful as an appliance: minimal C++ boilerplate, instant UI updates without a cloud round-trip, and OTA and logs in the same page you already opened to look at sensors.

Source: github.com/aaronbeckmann/ESP-DashboardPlus

---

5. public-apis — the collective list of free APIs

public-apis/public-apis (472k stars, 52k forks, 5,182 commits, MIT) remains the canonical directory of free public APIs.

It is neither a product nor a service — it is a curated index, maintained by pull request, that collects endpoints requiring no commercial key or offering a free tier, organized by category. That is its utility: when you need a quick data source for a prototype, test harness, or classroom exercise, this is where you look first before wiring credentials or scraping.

Longevity is the signal. A repo with this star count and commit cadence survives because every generation of builders needs the same list, and every API needs discoverability. Keep a local clone if you depend on it; entries change as providers do.

Source: github.com/public-apis/public-apis

---

Five finds, one record. Each note above will also post individually with a link back here.