This week in one line
Approaches that improve an agent's runtime environment without retraining the model showed up alongside Korean organizations building agents in house or plugging in Korean-made models.
We've gathered highlights from this week's Hypemarc news digest (2026-W43) to explore together with fellow builders. Some sources are Korean-language publications.
Seven news items this week
1. Google releases Harness², a framework that improves its own runtime
Google released Harness² as open source. It leaves model weights alone and instead improves the environment around the model: instructions, tool usage, memory, and task procedures. According to the report, a single round of improvement raised scores on three benchmarks.
Enterprise AX angle: Retraining is not the only lever for agent performance, so reviewing the prompts and tool setup of agents already in production is a reasonable first step.
Source: AI Times (Korean)
2. GitHub Copilot retires some models as of October 19
GitHub deprecated some Copilot models as of October 19, 2026. The change covers most experiences, including Copilot Chat, inline edits, agent mode, and code completion. The notice lists which models are affected.
Enterprise AX angle: Teams on Copilot should check that their settings and automation scripts do not pin a retired model.
Source: GitHub Changelog
3. Anthropic expands internal access for independent evaluators
Anthropic said it gave six people from the independent evaluator METR employee-level access to internal tools and systems. It is also working with SecureBio and Accenture on evaluations, and METR publishes its risk assessment results.
Enterprise AX angle: How a model provider is evaluated by outside parties can be added to security and compliance reviews when choosing a vendor.
Source: Anthropic
4. OpenAI improves coding in low-resource programming languages
OpenAI announced better ChatGPT coding performance in programming languages with few users. It also introduced new evaluations for software engineering, SQLite, and COBOL migration.
Enterprise AX angle: Financial and public sector teams with legacy COBOL can start by comparing the published evaluation tasks with their own migration work.
Source: OpenAI
5. Sendbird and Upstage bring a Korean LLM to customer service agents
Sendbird and Upstage signed an MOU. Upstage's LLM Solar will be applied to Sendbird's agent platform Delight.ai, with response speed, task accuracy, and customer satisfaction verified in stages.
Enterprise AX angle: Teams evaluating a domestic model can borrow the idea of fixing staged verification metrics up front, as this partnership does.
Source: AI Times (Korean)
6. Korea's Public Procurement Service has staff build AI agents themselves
Korea's Public Procurement Service is reported to be pursuing administrative innovation by having its own staff build AI agents instead of outsourcing development.
Enterprise AX angle: Having business staff build agents depends on training and operating rules being in place, so companies considering a similar approach should prepare both together.
Source: The Financial News (Korean)
7. Salesforce brings its "Partner FDE" agent talent program to Korea
Salesforce is bringing its "Partner FDE" program, which trains AI agent specialists, to Korea.
Enterprise AX angle: Companies looking to outside partners for agent builders gain one more option; checking a partner's actual build experience still matters.
Source: Chosun Biz (Korean)
Five trending GitHub repositories
We've picked repositories whose stars grew quickly this week. Repositories newly added to the resource library are gathered separately below.
anomalyco/opencode
The repository has around 218,000 stars.
What: An open source AI coding agent. Why it stands out: Useful for teams comparing an open source option alongside commercial coding agents. License: MIT.
mattpocock/skills
The repository has around 298,000 stars.
What: The author's everyday skills for coding agents, kept small and easy to adapt, with install instructions for Claude Code and Codex. Why it stands out: It favors composing small skills over a framework that owns the whole process, which suits teams building up their own rules. License: MIT.
feder-cr/git-ai
The repository has around 2,500 stars.
What: A version control system designed for coding agents. It is an independent implementation of the design Anthropic described publicly for its internal system Scion, and states that it is not affiliated with Anthropic.
Why it stands out: It targets setups where many agents keep landing short tasks on the same branches.
License: Per its LICENSE file, the ee/ directory has a separate license and everything else is MIT. GitHub could not detect the license automatically (NOASSERTION), so read the file before adopting it.
obra/superpowers
The repository has around 302,000 stars.
What: A framework that packages a development methodology as skills, so coding agents work through requirement checks, design, planning, and subagent-driven development in order. It installs into several coding agent tools. Why it stands out: We also shared it in this week's news digest. Useful for teams that want one development process across several agent tools. License: MIT.
Leonxlnx/earshot
The repository has around 1,600 stars.
What: A tool that lets coding agents analyze and make sound. It runs locally on CPU and works as a CLI, a Python API, and an MCP server. Why it stands out: It checks the audio in videos and demos that agents produce, which ties into marketing content automation. License: MIT.
This week's resources and videos
Five new resource entries: We reviewed and added five repositories to the AI resource library this week. Each entry notes what it is and what to check before adopting it.
- deepseek-ai/deepseek-harness: a plugin-centered agent harness
- openclaw/openclaw: an AI assistant you run on your own devices and connect to messaging channels
- browser-use/jev-ultrafast: a web agent that chooses actions on a page
- DietrichGebert/ponytail: coding agent skills that steer toward reusing existing code first
- affaan-m/ECC: skills, rules, and hooks for coding agents
Two videos:
- Cooking with Codex: Ep. 1 (OpenAI): Codex cloud tasks that keep running after you close your laptop, plus voice input, agent management, and conversation forking in the Codex CLI.
- Thomson Reuters on Building Trust in Legal AI (OpenAI): Thomson Reuters connecting its legal products to ChatGPT Enterprise, with an emphasis on output lawyers can verify and cite.
Share your take on this week's news in the community.
Wrapping up
The Hypemarc Builder Community will keep sharing the AI news and open source projects we explore each week. We'll continue gathering useful resources to share with fellow builders.
- Open chat: AI Builders
- Adoption advice and inquiries: Contact form
- Get the next post: Blog RSS · AI resource library