Releases·Tool Release
DeepSeek harness 0.2 adds Optional Bundles, async Question Mode, Windows Sandbox recovery, and official desktop clients
DeepSeek released harness 0.2, reorganizing the agent runtime around a Profile + Bundle model and adding experimental async user interaction. Windows and macOS desktop clients are now official, with account login hinting at paid plans. The release focuses on composability, permission boundaries, and recovery paths for production use.
88 engagement1 sourceThu, Oct 1, 2026, 07:59 AM
AI Overview
- DeepSeek harness 0.2 moves Schedule reminders out of defaults into an Optional Bundle, the first systematic use of the Profile + Bundle model for official features.
- A new Windows Sandbox permission-diagnosis skill detects Access Denied causes and performs backed-up, recoverable permission fixes after user authorization.
- Experimental Async Question Mode lets the Agent keep working after 'ask the user' timeouts instead of blocking synchronously on replies.
- DeepSeek-account sessions now run Web Search without an extra API key, with a updated third-party model catalog, fuzzy search, and keyboard navigation for model lists.
- Official Windows and macOS clients released (Linux unsupported), with account login suggesting paid plans may come soon.
Related
Products·launched
Fulmar: native macOS app wrapping DeepSeek Harness agent runtime with access, model-routing and recovery controls
Fulmar is a native Mac app that wraps the DeepSeek Harness (DSH) agent runtime, which handles conversations, tools and tasks. It adds management controls the raw runtime lacks: choosing local or cloud model routes with explicit consent, Keychain credential storage, workspace/tool access limits, and recovery checkpoints. It targets developers running file- and command-capable agents on macOS who want oversight and safety.
Fulmar is a native Mac app that wraps the DeepSeek Harness (DSH) agent runtime, which handles conversations, tools and tasks. It adds management controls the raw runtime lacks: choosing local or cloud model routes with explicit consent, Keychain credential storage, workspace/tool access limits, and recovery checkpoints. It targets developers running file- and command-capable agents on macOS who want oversight and safety.
10 engagement1 sourceredditSun, Sep 13, 2026, 10:30 AM
Developments·Paper
Arxiv paper analyzes architectural convergence in three LLM agent harnesses
A new arxiv paper examines three open coding-agent harnesses—LangChain's deepagents, Earendil's pi, and DeepSeek's dsh—highlighting architectural convergence despite differing design philosophies. The study argues that the harness layer, not the underlying model, is the primary constraint on agent behavior, offering insights for practitioners building autonomous agents.
A new arxiv paper examines three open coding-agent harnesses—LangChain's deepagents, Earendil's pi, and DeepSeek's dsh—highlighting architectural convergence despite differing design philosophies. The study argues that the harness layer, not the underlying model, is the primary constraint on agent behavior, offering insights for practitioners building autonomous agents.
0 engagement1 sourcearxivTue, Aug 25, 2026, 01:26 AM
News·Incidentfeatured
GitHub repo documents 2026 China LLM distillation scandal involving encrypted CoT extraction from Claude
A GitHub repository titled 'china-llms-anecdote-0721' compiles insider information about a major incident in China's LLM industry from April to July 2026. It details how multiple major AI labs, including Zhipu GLM, Tencent Hunyuan, Moonshot Kimi, Qwen, and DeepSeek, extracted and distilled Anthropic's encrypted chain-of-thought (CoT) from Claude, and covers the subsequent controversy over Kimi K3's architecture and benchmark manipulation.
A GitHub repository titled 'china-llms-anecdote-0721' compiles insider information about a major incident in China's LLM industry from April to July 2026. It details how multiple major AI labs, including Zhipu GLM, Tencent Hunyuan, Moonshot Kimi, Qwen, and DeepSeek, extracted and distilled Anthropic's encrypted chain-of-thought (CoT) from Claude, and covers the subsequent controversy over Kimi K3's architecture and benchmark manipulation.
021153 engagement3 sourcesgithub, rssTue, Jul 21, 2026, 01:30 AM
Releases·Model Releasefeatured
DeepSeek-V4-Flash-0731 release sparks market crash fears
DeepSeek released V4-Flash-0731, a new model that outperforms GLM 5.2 at the same cost as its predecessor. The release has triggered concerns about another market crash, reflecting the disruptive impact of DeepSeek's competitive pricing and performance.
DeepSeek released V4-Flash-0731, a new model that outperforms GLM 5.2 at the same cost as its predecessor. The release has triggered concerns about another market crash, reflecting the disruptive impact of DeepSeek's competitive pricing and performance.
785 engagement1 sourceredditFri, Jul 31, 2026, 06:57 AM
Releases·Tool Release
Cline Desktop v0.0.38 adds Linux ARM support and defaults OpenAI providers to GPT-6.1 Sol
Cline shipped Desktop v0.0.38 on GitHub, adding Linux ARM64 packages, cloud session persistence, and a range of reliability fixes. The release also refreshes the default model catalog across providers, making GPT-6.1 Sol the default for OpenAI and most gateways.
Cline shipped Desktop v0.0.38 on GitHub, adding Linux ARM64 packages, cloud session persistence, and a range of reliability fixes. The release also refreshes the default model catalog across providers, making GPT-6.1 Sol the default for OpenAI and most gateways.
0 engagement1 sourcegithubWed, Sep 30, 2026, 03:16 AM