Lovable: we deleted almost all our agent code
Lovable's lessons from 100k PRs: choices that once helped can hold you back, so the team keeps rebuilding how the agent works.
AI Daily Recap
17 articles · 4 categories
The finishable daily brief
What happened in AI — Nov 8, 2026
Sunday, Nov 8, 2026
17 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Lovable deleted almost all its agent code after 100k PRs.
- Anthropic reportedly cut over 80% of Claude Code's system prompt with no measurable eval loss.
- Agent memory lacks a shared format; the OpenTelemetry analogy is gaining ground.
- ARTEX-driven campaigns stole 190k gaming credentials for under $400 in model costs.
- Uncensored models in agent loops: most fail about half of multi-step tasks.
Builders are shrinking their agent stacks: Lovable deleted most of its agent code and one engineer dropped all agent skills, citing Anthropic's 80%+ Claude Code prompt cut with no measurable eval loss.
The infrastructure gap is memory: every coding tool stores it differently, and several projects propose a shared format or an inspectable store.
Prefer to watch?
LLM Digest Daily Explainer: Nov 8, 2026
Agent harnesses: less scaffolding, more deletion 5 items
Builders report pruning agent code and skills as models improve, and the lighter harnesses they keep are small and inspectable.
Author deletes all their agent skills
Cites Anthropic cutting over 80% of Claude Code's system prompt with no measurable loss on its own evals, and argues skills can over-constrain the agent.
psi: a 6.4 MB coding agent in ANSI C and Lua
MIT-licensed core in C with the agent loop, tools and providers in Lua; works with Anthropic, Codex, OpenRouter or Ollama.
My Local Coding Agent
A 25-year programmer describes a local agent setup and says LLMs in agent harnesses became capable of full features in January 2026.
Designing an Always-On Agent
Argues today's agents do nothing until messaged and explores what a continuously running, environment-aware agent needs.
Agent memory and context become shared infrastructure 5 items
Memory is fragmented per tool, and new projects try a shared format or an inspectable filesystem-style store.
Agent memory needs its OpenTelemetry moment
Each coding tool stores memory in its own shape and none can read the others; the post surveys existing formats and sketches a shared spec.
OpenViking: a context database for agents
Open-source store that exposes knowledge, memory and skills as a viking:// virtual filesystem agents navigate with ls, tree, read and grep.
RepoMagpie: your coding agent remembers repos you've tried
Keeps your verdicts on libraries, CLIs and MCP servers so the agent checks them before installing a tool you already ruled out.
Google's Elephant-Goldfish Model plugin
Splits work in two: Elephant keeps project context and drafts the spec; Goldfish reviewers read it without the prior conversation to expose blind spots.
aiport: a local LLM gateway
Go gateway that routes Claude Code, Codex, Ollama and cloud models through one endpoint and shows which app called what.
Security and reliability of agents under test 4 items
Offensive agent use is cheap and measurable, while new work targets model backdoors and agent-loop failure rates.
Check Point: AI agents stole 190k gaming credentials for under $400
Analysis of ARTEX, an open-source LLM-driven pentesting platform, and how cybercrime campaigns used it at low model cost.
What Can AI Hack? Inside Frontier Labs' Tests
Explains how to read system-card cyber evals, including OpenAI rating GPT-6 Astra its first Critical-level cyber model in September 2026.
Uncensored LLMs in agent loops: 31 models, 5 tasks
Most models still fail about half the time on multi-step tasks; models built on weaker bases loop almost twice as often.
NC State: new techniques to prevent AI backdoor attacks
Three techniques to block embedded triggers, plus a proposal to reuse backdoor methods as a security layer for users.
Inference and fine-tuning off the CUDA path 2 items
Two items show serving and tuning stacks moving beyond Nvidia and Python.
VindexLLM: GGUF inference in Delphi on Vulkan
Runs full transformer inference on the GPU through Vulkan compute shaders, with no Python or CUDA, needing only vulkan-1.dll.
Huawei Ascend adapts Kimi K3.1 Pro for LoRA fine-tuning
Fine-tuning runs on a 128-card Atlas 950 SuperPoD, per Pandaily.
You are caught up for this edition