Newsletter 20 August 2026
How this is put together. This page is filled automatically every day from public sources (X, Hacker News, GitHub, YouTube). The headlines come from the source, the Dutch sentence underneath is machine-written. Click through to the source before you base anything on it.
Replit
- 【Replit「Free Mode」提供開始のお知らせ】 本日、Replit は Agent の新しい使い方となる Free Mode を発表しました。OpenAI との共同発表で、基盤には GPT-5.6 Luna を採用しています。あわせて、チャットから成果物までを一続きで扱
- I was shocked by how much Free Mode is capable of, actually gasped in internal staff testing. I remember the early days, and the app I was a
- 🚨 Replit launches FREE MODE, so you stop rationing your prompts. This new mode, powered by new models, is expected to provide approximately
- Senior Product Manager
Claude Code & agents
- AI Coding Tools Ranked by Use Case 🟢 Best for Fast Prototyping ⚡ Lovable ⚡ Bolt ⚡ v0 🔵 Best for AI Coding 🤖 Claude Code 🤖 Codex 🤖 Cursor 🟣 B
- chore(deps): update Claude Agent SDK to 0.3.235
- Cisco’s Claude Code write-up shows why persistent memory changes the agent security model: poisoned state can survive across projects, sessi
- [BUG] 2.1.237 tagged `latest` with its linux-x64, win32-x64 and linux-x64-musl native packages never published - installs land on a dead 500-byte stub
Cloudflare OS & self-hosted
- AI agent reliability shouldn’t require stitching five tools together. Future AGI is an open-source, self-hostable platform for evaluating, o
- Yes, the framing is useful. The self-hosted vs managed proxy split (LiteLLM for cost/privacy control versus Portkey/Helicone/Braintrust for
- Langfuse v4 is live on Langfuse cloud and GA for self-hosters. v4 is a re-architecture of our data model. It is more performant and enables
- Our self-hosted Langfuse stack stored 14GB on disk, only 1.6GB of it was actual observation data
Agent tooling & MCP
- That gap has a name: MCP, the open standard for plugging an agent into your orders, billing and tickets. We ship one in ChatterMate, so I wr
- A Google engineer just dropped a 421-page guide on Agentic AI. 🤯 Inside: → Prompting, routing & memory → MCP & multi-agent systems → Reasoni
- [P1][AI/Security] Agent / MCP Tool Authorization / Approval / Execution Evidence
- Black Hat Asia 2026 | Remote Server, Local Root. Welcome to MCP.
Models & pricing
- chore(pricing): document OpenAI Fast mode/Flex tier coverage gap
- Most AI cost overruns are not a pricing problem. They are a routing problem. A classification task that could run on a small model runs on a
- New benchmark just dropped. Not another "can it solve the end of the world" exam. Not another biology or philosophy olympiad. This one answe
- feat(llm): add moonshot/kimi-k3 to model prices and context window map