Introduction

From the data collected on July 25, we've picked six stories, focusing on topics not covered in the digest articles already published today: the trade-off between API cost and performance, "harness design" in AI development, the reality of bringing agents to legacy code, and hands-on insights from engineers in Japan — perspectives closer to the front lines.

---

Tencent HY3: Sonnet-Class Performance + Reasoning, Completely Free on OpenRouter

Writing on Qiita, lumichy introduces Tencent's new model, **HY3**. It's drawing attention for offering **Sonnet-class performance and reasoning capabilities completely free via OpenRouter**, addressing the trade-off between API cost and reasoning ability — the biggest bottleneck in AI agent development.

Stick with current OpenAI or Anthropic frontier models, and an agent that repeatedly calls tools can burn through millions of tokens in no time; running into so-called "token bankruptcy" is far from rare. HY3's approach broadens your cost options without fully sacrificing performance, so it looks highly practical for PoCs and validation work. That said, the sustainability of the free offering, rate limits, and stability in production will presumably need separate verification at this point.

TS Compiler Knowledge Graph: An Approach That Cuts AI Tokens by About 90%

A **Knowledge Graph** package for the TypeScript Compiler, published on GitHub by samchon, is making the rounds on Hacker News. It structures TypeScript type information and AST as a knowledge graph and hands the LLM only the context it needs instead of "raw source code", claiming to **cut AI token consumption by about 90%** while maintaining accuracy.

Throwing a massive codebase wholesale at an LLM is nearing its limits in both accuracy and cost, and "structuring and extracting only the necessary context" is a challenge that major agents like Cursor and Claude Code are all working on. While it's positioned as a TypeScript-specific knowledge graph, it's also worth watching from the angles of porting it to other languages and applying it to internal codebases.

An Experiment Bringing Agile AI Development to a Million-Line Legacy SaaS

A case posted by thegreatkahuna on Hacker News's Ask HN is rich with implications for thinking through the realities of AI agent development. Against a **SaaS with over 1 million lines of C# and React code and a 15-year history**, they had a Claude-based agent build an MVP with no engineers on hand, reportedly completing 13K lines of feature code plus an equivalent amount of tests in roughly two weeks.

Particularly interesting are the process-design choices. The agent owned the PRD → architecture document → Epic breakdown → Story decomposition flow, and PR reviews were run agent-end-to-end. The poster cautions, however, that "when Claude reviews code written by Claude, it misses the same errors," and reports that **cross-review with a second model (Codex) caught substantially more issues**. That's a valuable firsthand data point on the limits of a coder-reviewer setup built on a single model. The heart of the post is a practical question — "how do we get the code into a state we can trust before the engineer review in August?" — and further answers from the community are expected.

Instavm: Lifting Small-LLM Capabilities with "Code Mode"

Instavm's blog post "How Code Mode Can Help Smaller LLM Models" was featured on Hacker News. Across multiple benchmarks, it shows that so-called "code mode" — having the model express its reasoning process in code rather than natural language — **works effectively for smaller models too**, not just large ones.

As more teams run small models in local inference and edge environments, codifying CoT (Chain-of-Thought) to achieve both reasoning stability and accuracy is a practical direction. The applicability is broad, from suppressing hallucinations to playing well with structured output. For developers betting on local operation, this is an approach worth keeping in mind as a tuning candidate you can try right away.

From AI Development Trenches in Japan: SDD Harnesses, the Honesty of a No-Go, and Code Review Automation

Today on Qiita, three posts digging into the practice of AI development went live side by side.

**1. Making AI implement "to spec" with an SDD harness (NeoSoleil)**

The latter half of a series built around iMAS Chara Match, an iMAS character diagnostic tool, describing a **harness that structurally prevents** problems like AI ignoring the spec and building whatever it likes, skipping tests, or slipping past review. It's a hands-on example of assembling a spec-driven development (SDD) framework on top of Claude Code, and what sets it apart is the stance of treating process design itself as a deliverable rather than being just another "we let AI build it" story.

**2. The honesty of choosing a "temporary No-Go" after seven days (boraemon2000)**

A record of the decision to deliberately call a "temporary No-Go" on Day 7 of the RC (release candidate) validation week, on the grounds that **even with tests passing, uncertainties remained in the unattended every-morning operation**. What makes it interesting is the reframing: the author says their past self would have taken it as a failure and been disappointed, but they came to see it as "honesty as an operating OS". How to pair Go/No-Go decision criteria with automation is a recurring theme in the design of AI operations.

**3. A supplement on automating AI code review (bug-mi-dev)**

A companion piece to the video "Automating Code Review with AI". It reports an average saving of 20 minutes per PR while carrying the message that "this is no longer an era of doing everything by hand". Its position is a measured one, though: it keeps the frame of "letting AI handle the initial checks" without breaking the structure in which humans own the final judgment.

What all three share is the attitude of making the line — what you delegate to AI versus what humans vouch for — explicit through process. The philosophy of not treating the automation ratio itself as the goal, and of adopting AI only as far as reviewability and operational safety can be maintained, looks set to become the baseline for human-AI collaborative development going forward.

US Department of Energy Announces First "AI Genesis Mission" Projects

US Secretary of Energy Chris Wright announced the first **AI Genesis Mission** project selections. The entry shared by donutloop on Hacker News offers few details, but the initiative is positioned as a framework to accelerate AI adoption in the energy sector.

Judging from the name "Genesis Mission", it appears to be a policy package bundling new infrastructure, data, and talent to drive AI adoption, but public information is limited at this point. We'll be keeping an eye on future official announcements, the substance of the selected projects, and spillover effects into the civilian sector.

---