Coding Agents & Loop Engineering

• The creator of Claude Code stopped prompting — Boris's 14-step system: loops prompt Claude, humans write loops; triggers, execution skills, goal clarity, and verification are the actual job now.
• $50k loop failure modes — three root causes firms charge fortunes to fix: stopping on terminal message not task completion, context rot compounding bad decisions, and missing maker-check passes.
• I spent 2 years on AI engineering philosophy → 4 pre-build gates — distilled framework for gating agents before they build: the gates matter more than the model.
• Context windows vs task handoffs — the real debate: bigger windows delay the rot but structured handoffs fix the underlying problem; longer context is a crutch.
• Keep your agent's context window sharp — practical tactics for pruning context rot before it degrades agent decision quality mid-task.
• SQLite → Zig port: 90/102 translation units done, 919M tokens in — real-world proof that Claude Code can sustain long mechanistic translation tasks against an authoritative test suite.
• Production iOS + Android app shipped with Claude as primary dev tool — honest postmortem: what worked (scaffolding, boilerplate, iteration) and where the agent still falls down in real production work.
• AI coding bottleneck is no longer coding — after 2 months with agents, the constraint shifts to spec clarity, review speed, and knowing what to build — not writing the code.
• ~6 months until agents move off laptops — Theo's call: cloud-hosted persistent agents replace the local Claude Code session as the standard dev environment.
• Wayfinder Router: deterministic local/cloud LLM routing — routes prompts to local or hosted models based on prompt shape (length, structure) with no model call to decide — fully offline, calibratable on your own data.
• DSpark: speculative decoding accelerates LLM inference — DeepSeek's production speculative decoding paper; explains the throughput edge behind their pricing versus competitors' capex bets.
• OpenSpec: spec-driven development for AI coding agents — artifact-guided workflow that keeps specs fluid and iterative, designed for brownfield codebases not just greenfield; /opsx:propose to start.
• claude-howto: visual example-driven Claude Code guide — copy-paste templates covering basic usage through multi-agent orchestration, hooks, skills, and MCP servers with a structured learning path.
• Why are Claude Code skill files online all pointless? — useful meta: the quality gap between public skill files and real production-grade ones is large; most published skills are demo-tier.
• ThePrimeagen on MCP — "the people who say coding is solved also created MCP" — pithy reminder that protocol complexity is a sign the problem isn't solved.

CLI / TUI

• herdr: agent multiplexer in your terminal — workspaces, tabs, panes, mouse-native split/drag, detach-and-reattach with agents running; no Electron, no GUI wrapper — pure terminal agent orchestration.
• ccgram v4.3.0: control Claude Code from Telegram via tmux/herdr — mobile remote control for Claude Code sessions; useful for async loop monitoring without staying at a desk.

OSS Models & Local Agents

• Ornith 35B MoE with self-scaffolding RL — runs on a single desktop at near full precision, 3M+ token context, self-verifies tool calls mid-task; beats Qwen3.6 on coding benchmarks.
• Ornith 397B benchmarks near Claude Opus 3 — Chinese open-source; 35B for local use, 397B for hosted; another data point in the open-weight capability curve compressing fast.
• Gemma 4 12B Agentic Fable5 GGUF — locally-run quantised model tuned for coding, terminal tasks, and agentic tool use; 206k downloads signals real adoption.
• opencode: open-source coding agent — curl | bash install, npm/bun/scoop support; positions as the open alternative to Claude Code and Codex for terminal-native coding.
• Superpowers composable skills framework for Claude Code — gating skills that force planning and test-writing before code generation; solves the "agent jumps straight to files" failure mode.

Agentic SaaS & Memory

• EverOS: portable memory layer across agents — local-first Python runtime storing conversations, files, and agent trajectories as Markdown with SQLite + LanceDB indexing; one memory layer shared across Claude Code, Codex, and others.
• Folio AI: Claude for PowerPoint — AI slide generation on top of Claude; narrow vertical SaaS wrapping the model for document-heavy enterprise workflows.