This week in one sentence
Team-managed agent tools and contact center results arrived alongside an incident report that puts customer data isolation in inference services back on the adoption checklist.
For 2026-W44, we're exploring Cursor Agent SDK, Copilot's shared workflows and context management, and KT's AI contact center. We've also included NVIDIA's local model and Anthropic's proposed IPO.
For builders evaluating adoption, operating conditions deserve attention alongside features. Each story points to questions about who can change an automation, where costs land, and how customer data stays separate.
Seven news stories this week
1. Cursor Agent SDK brings team-specific agents into code
Cursor released the Agent SDK used for its own agents, including Bugbot. It pairs Markdown instructions with TypeScript tools and supports evaluations and event replay. Agents can run on Cursor's infrastructure or your own.
Enterprise AI perspective: For recurring tasks such as PR review, put instructions, tools, and evaluations through code review and decide who approves changes to business policies.
Source: Cursor
2. KT reports improved resolution and handling time in its AI contact center
According to AI Times KR, KT says its voicebot handles an average of 1.2 million consultations monthly, with resolution rising from 77% to 81%. Average handling time fell 11% after introducing an assistant for human agents. The voicebot also handles follow-up work, including fault intake and remote remediation.
Enterprise AI perspective: Evaluate customer service automation using resolution, handling time, and the conditions for handing work to a person, alongside response volume.
Source: AI Times KR
3. Cloudflare reports cross-customer data exposure in GLM-5.3
Cloudflare disclosed that 155 GLM-5.3 responses on Workers AI contained other customers' request data between August 27 and September 9. A KV cache configuration mismatch and a SGLang bug affected inputs above 681,280 tokens. Cloudflare says it fixed both issues and other models were unaffected.
Enterprise AI perspective: When evaluating inference for long documents, ask about cache isolation and error handling for oversized requests as well as maximum input length.
Source: Cloudflare
4. NVIDIA introduces Nemotron 3.5 Bolt for roughly 4GB of memory
AI Times reports that NVIDIA released Nemotron 3.5 Bolt E4B 4GB for local agents, robots, and related uses. It supports tool calls, switchable reasoning, and a context window of up to 64,000 tokens. The report also notes weaker results than a comparison model on some tool-use evaluations.
Enterprise AI perspective: For local deployment, test tool-call accuracy and response latency on your own tasks alongside memory requirements.
Source: AI Times
5. Copilot Automations expands to shared team workflows
GitHub is rolling out shared automations combining prompts, triggers, and allowed tools. Repository readers can view them, writers can edit them, and AI credits are billed to the repository owner. Organization repositories require an administrator to enable the relevant usage policy.
Enterprise AI perspective: Review repository permissions, organization billing policy, and trigger change history so that an automation can keep operating when its maintainer changes.
Source: GitHub Changelog
6. GitHub intelligence connects agents with team conventions and history
GitHub introduced GitHub intelligence around memory, context improvement through dreaming, and code search. Copilot Memory is generally available on paid plans; dreaming is an opt-in public preview. Changes to repository skills and instruction files are proposed through pull requests.
Enterprise AI perspective: Define which repositories and context agents may access, and review proposed instruction changes before adopting them.
Source: GitHub Changelog
7. Anthropic files an S-1 for a proposed IPO
Anthropic announced an S-1 registration filing with the U.S. Securities and Exchange Commission for a proposed initial public offering. It plans to list on Nasdaq under “ANT.” The share count and price range remain undetermined.
Enterprise AI perspective: Teams considering long-term model supplier contracts can incorporate future financial and business risk disclosures into vendor evaluation.
Source: Anthropic
Five GitHub projects to explore
From repositories gaining stars quickly, we've selected five that offer useful approaches to development workflows and agent infrastructure.
DietrichGebert/ponytail
Around 181,000 stars.
What: A coding agent skill that encourages examining existing code, standard libraries, and platform features before writing new code. Why it matters: Teams can use it as a reference for agent instructions that favor reuse of internal code and built-in capabilities. License: MIT.
deepseek-ai/deepseek-harness
Around 266,000 stars.
What: An open source agent runtime organized around plugins, with a web UI. Why it matters: It offers an approach to assembling capabilities as plugins. The project is a developer preview with breaking changes expected, so review its version and safety guidance before adoption. License: MIT. Third-party dependency notices are provided in THIRD_PARTY_NOTICES.md.
mattpocock/skills
Around 307,000 stars.
What: Small, adaptable, composable coding agent skills for clarifying requirements and carrying out engineering work. Why it matters: Builders can select the skills that fit their existing development process and adapt the content to their team. License: MIT.
affaan-m/ECC
Around 293,000 stars.
What: Skills, rules, and hooks for coding agent workflows, including planning, verification, review, and memory. Why it matters: It provides operating conventions to compare across agent tools. Define the installation scope first, as overlapping installation methods can duplicate skills and hooks. License: MIT.
deepseek-ai/DeepInfer
Around 24,000 stars.
What: An inference framework for DeepSeek models. The public release supports offline inference and structured outputs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Why it matters: Teams evaluating their own inference infrastructure can explore its execution design. A complete service still needs components such as a KV cache system and load balancing. License: Apache-2.0.
This week's resources and videos
There are no new resource library entries this week. Browse the AI resource library for agent tools and adoption considerations.
Three videos to explore:
- Trying unlimited local AI on a ₩15 million MacBook (Yoon Automation, in Korean): A look at writing, coding, and audio tasks with local models, comparing speed, power use, and cost.
- Agents on longer tasks need more than one loop (DeepLearningAI): An introduction to a course covering agent loops alongside verification, event-driven, and trace-based improvement loops.
- AI shared my secrets: Cloudflare's incident and why NVIDIA built a DPU (안될공학, in Korean): An explanation of inference cache failures and customer data isolation, to explore alongside Cloudflare's report above.
Share your thoughts on the week's news with 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