We Tested Coding Agent Harnesses. Thinking Harder Beat All of Them
Across 100 tasks, a 37-skill, 17-agent harness cost 8× more and scored below vanilla Claude Code with more thinking.
AI Daily Recap
32 articles · 5 categories
The finishable daily brief
What happened in AI — Nov 4, 2026
Wednesday, Nov 4, 2026
32 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Extended thinking beat a 37-skill, 17-agent harness that cost 8× more.
- NVIDIA: frontier agents lose little accuracy at 16K context with compaction; long context costs up to 7× more per token.
- Cleric's ~40-agent rewrite made evals 5× faster and 100× cheaper.
- Kimi K3.1 Pro (2.8T MoE, open weights) had day-0 vLLM support.
- Glow Labs: coding agents are installing deprecated MCP servers with known flaws.
- H200 loses money fully busy; B300 profits at 35% utilization.
Heavier agent scaffolding lost to simpler setups: a 100-task benchmark found Claude Code with more thinking beat a 37-skill, 17-agent harness that cost 8× more.
NVIDIA's result that 16K contexts with compaction hold accuracy points the same way: cut context and orchestration, not capability. Kimi K3.1 Pro's day-0 vLLM support shows open-weight serving catching up fast.
Prefer to watch?
LLM Digest Daily Explainer: Nov 4, 2026
Agent harnesses: less scaffolding, durable memory 5 items
A 100-task benchmark found extended thinking beat elaborate harnesses, while three new tools move agent memory into portable stores.
Renor: one memory for all your AI agents
Apache-2.0 Markdown-notes memory recalled over MCP, with clients for Claude Code, Codex, Cursor, Gemini and VS Code.
Engraphis: local-first, bi-temporal memory for coding agents
Inspectable, self-hostable memory with code-aware recall and bi-temporal history, served over MCP.
Mentat: one memory for all my AI tools
Linked notes in a private git repo, served to Claude and OpenCode from a small Cloudflare Worker.
DeepSeek Open-Sources Super-Agent, a Coding Agent That Works From Team Chat
DeepSeek released a coding agent driven from team chat, as an open-source project.
Evals and context: cheaper runs, shorter windows 3 items
Two results cut the cost of agent operations: Cleric's 40-agent eval rewrite and NVIDIA's finding that 16K contexts with compaction hold accuracy.
How we made evals 100× cheaper with multi-agent architecture
Cleric split one investigation agent into ~40 loosely coupled agents, running evals 5× faster for cents.
Nvidia Study: AI Agents Rarely Need Long Context Windows
Claude Opus 4.5 and GPT 5.6 Sol lose little accuracy at 16K with frequent compaction; long-context serving costs up to 7× more per token.
Evalship: find the flaws in your LLM evals
Watches every pull request and auto-optimizes prompts against your existing eval suite.
Inference economics and open-weight serving 5 items
Kimi K3.1 Pro had day-0 serving support, while two measurements show gateway language and GPU choice matter more than engine choice.
vLLM Delivers Day-0 Support for Kimi K3.1 Pro
Moonshot's 2.8T-parameter (104B active) open-weight MoE runs on NVIDIA, AMD, TPU and Alibaba PPU from launch.
Verda Inference, in Preview with Kimi K3.1 Pro
Verda adds managed model-as-a-service inference for open-weight frontier models, launching with Kimi K3.1 Pro.
Is hosting an open-source inference engine profitable?
H200 loses money even fully busy; B300 turns a profit at 35% utilization, so the GPU matters more than the engine.
Go vs. Python for an LLM Gateway: What We Measured
Reproducible benchmark of KTMS (Go) against LiteLLM Proxy and eight other gateways on latency, TTFT and memory.
How big is the open model economy today?
Interconnects uses its open-model dashboard to show adoption metrics are noisy and need an informed read.
Agent security and identity 4 items
Coding agents are installing retired MCP servers on their own, and NIST and Cloudflare each publish controls for identifying and governing agents.
AI coding agents are installing abandoned MCP servers with known security issues
Glow Labs' "MCPhantom" pattern: agents pick deprecated MCP servers that no longer receive security fixes.
Design Tenets for Agent Identity Management Systems
NIST's second post on agent identity, authentication and authorization sets design tenets for identity systems.
See the agents, set the terms: new visibility and control for personal agents
Cloudflare adds visibility into which personal agents visit a site and lets owners set terms for them.
Daybreak Red is now available in AWS GovCloud (US-East)
GPT-5.6 Cyber reaches approved customers through Amazon Bedrock for authorized cybersecurity work.
Distribution and spend: Qwen on iPhone, Claude Code economics 3 items
Distribution and spend drive platform strategy: Apple's Qwen trial in China, OpenAI buying into Codex tooling, and Anthropic's Claude Code margins.
Apple Begins China Trial Of AI Tools Powered By Alibaba's Qwen
Apple's China beta links Qwen with Siri, Writing Tools and Visual Intelligence.
OpenAI to acquire Pierre Computer Company, Inc.
The acquisition expands Codex's tools for building, understanding and reviewing code.
Anthropic launched tools to show companies what their AI spending is buying
Claude Code reportedly hit an operating profit in Q2 with 900% net revenue retention.
You are caught up for this edition