After Testing DeepSeek Harness All Night: It Outperforms Claude Code in a Minecraft-like Way - 36Kr
An overnight hands-on test from 36Kr claims Harness edges out Claude Code on open-ended agentic coding tasks.
AI Daily Recap
26 articles · 5 categories
The finishable daily brief
What happened in AI — Aug 14, 2026
Friday, Aug 14, 2026
26 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- DeepSeek's open-source Harness gets its first hands-on reviews against Claude Code, with one overnight 36Kr test calling it the stronger agent.
- V4-Pro's higher API price than V4-Flash complicates DeepSeek's "better and cheaper" pitch.
- GitHub, AWS Bedrock AgentCore, and MongoDB each shipped ways to wire live infrastructure into coding agents today.
- Meta open-sourced Muse Glimmer, a 30B on-device agentic model under Apache 2.0.
- Anthropic detailed how Claude's upcoming text watermarking will work.
A day after DeepSeek open-sourced its plugin-based Harness alongside the pricier V4-Pro flagship, press on both sides of the Pacific ran it head-to-head against Claude Code — with one overnight test calling Harness the stronger agent, even as V4-Pro's API price undercuts DeepSeek's cheaper-than-Claude pitch.
Around that story, the agent-tooling stack kept filling in: GitHub, AWS, and MongoDB all shipped ways to plug live infrastructure into coding agents, while Meta and AMD pushed open-weight models further onto local hardware.
DeepSeek's Open Harness Gets Its First Reviews 4 items
A day after shipping the open-source Harness alongside V4-Pro, DeepSeek drew hands-on comparisons against Claude Code from press on both sides of the Pacific — with one head-to-head test calling Harness the stronger agent, even as V4-Pro's steeper API price complicates the cost pitch.
Against Claude Cowork, DeepSeek opens its open-source Harness to developers - TechNode
TechNode positions Harness as a direct, self-hostable alternative to Claude Code for developers wary of vendor lock-in.
DeepSeek Open-Sources the Missing Layer Between AI Models and Agents - HPCwire
HPCwire frames Harness's plugin architecture as the missing orchestration layer between raw models and working agents.
DeepSeek launches V4 Pro model with enhanced AI agent capabilities - news.cgtn.com
V4-Pro ships as DeepSeek's flagship model with agent-oriented upgrades, priced above the cheaper V4-Flash tier that first rattled Western labs.
The Agent-Tooling Stack Keeps Filling In 5 items
GitHub, AWS, and MongoDB each shipped ways to plug live infrastructure into coding agents today, while independent tools tackled the config-sprawl and research-budget problems that come with running several agents at once.
How to bring your software delivery workflow into GitHub with agent apps
Four GitHub agent apps now cover scoping, securing, rolling out, and shipping a feature without leaving GitHub.
Building agentic workflows with SageMaker AI and Bedrock AgentCore
AWS shows how to route a multi-agent workflow's specialized agents to whichever model fits each one best, including OpenAI-compatible SageMaker endpoints, via Bedrock AgentCore.
MongoDB Brings Live Operational Data to the Agentic Coding Stack - SMEStreet
MongoDB adds live operational data access to the agentic coding stack, giving coding agents a read on production state instead of stale schemas.
Agentstow: Canonical configs, fanned out to all your AI coding agents
Agentstow keeps one canonical config and fans it out to each AI coding agent's own format, aimed at the config-drift problem of running several agents side by side.
Show HN: Mole – Deep research agent for your terminal
Mole is a terminal-based deep research agent built to keep sources straight and stay within budget, problems its creator says plague typical agent research runs.
Evals, Benchmarks, and Inference Engineering 4 items
Today's technical deep dives focused on measuring what agents and inference stacks actually do: evaluating SRE agents against real incidents, benchmarking inference cost, cutting context bloat, and scheduling GPU verification work by confidence instead of brute force.
Evaluating AI SRE Agents in Production (OpenSRE) – Evaluation
OpenSRE lays out a framework for evaluating AI SRE agents against real production incidents rather than synthetic benchmarks.
Presentation: The Right 300 Tokens Beat 100k Noisy Ones: The Architecture of Context Engineering
Sadogursky and Debois argue coding agents fail from bloated, stuffed context windows, and lay out fixes including lazy-loaded skills and versioned context.
Adaptive Verification in vLLM: DSpark confidence-scheduled verification
vLLM's DSpark sizes its draft-verification budget from each request's confidence instead of verifying every drafted token, holding the throughput/latency frontier from batch size 1 to 256.
LLM Inference Benchmarking
DigitalOcean's benchmarking guide compares LLM inference cost and latency across serving configurations.
Open Weights Keep Spreading Past DeepSeek 3 items
Meta open-sourced a 30B on-device agentic model, Google's Gemini 3.7 Flash drew fresh attention, and AMD published a guide for running Qwen locally on its newest agentic PC silicon — a reminder that open weights are moving beyond DeepSeek this week.
Meta Open-Sources Muse Glimmer: A 30B Local Agentic Model Optimised for On-Device Execution
Muse Glimmer is a 30B open-weight model under Apache 2.0 built to run autonomous agents and complex tasks locally on consumer GPUs.
[AINews] Gemini 3.7 Flash brings GDM back to the forefront
Latent Space's roundup credits Gemini 3.7 Flash with putting Google DeepMind back in serious contention after a quiet stretch.
Run Qwen 3.8 27B on AMD Ryzen™ AI Max Agentic PCs and Radeon™ GPUs - AMD
AMD publishes a guide for running the 27B Qwen 3.8 model locally on its Ryzen AI Max chips and Radeon GPUs, aimed at on-device agent workloads.
Provenance and Security Controls Get More Granular 2 items
Anthropic detailed the mechanics behind Claude's coming text watermark, and Cloudflare shipped a one-click way to lock down internal, AI-generated apps — both aimed at the trust gap opened by AI-written code and text.
How Claude's text watermarking works
Future Claude models will embed a watermark in generated text so its likely AI origin can be checked later, a move Anthropic says several other major providers are also making.
Secure all your internal vibe-coded applications — in one click
Cloudflare Access for Workers lets teams attach an access policy directly to a Worker so it's enforced everywhere that Worker runs — routes, custom domains, and previews alike.
You are caught up for this edition