Dev Digest — March 7, 2026

🔥 HOT RELEASES

T3 Code — Open Source Coding Agent Manager Theo just dropped T3 Code, a GUI built on top of OpenAI's Codex CLI for managing multiple coding agents in parallel. It's fully open source, works with your existing Codex subscription, and is designed for devs who run lots of agents simultaneously. Community already comparing it to Codex App, Conductor, and Antigravity. Works on mobile too. 🔗 X: https://x.com/theo/status/2030071716530245800 💻 GitHub: https://github.com/t3dotgg/t3-code

GPT-5.4 Pro (xhigh) — Massive Benchmark Jump OpenAI's GPT-5.4 Pro achieved a 10-point gain on CritPt, the largest incremental gain ever tracked by Artificial Analysis. Comes with native computer operation and major cost reductions. However, reviewers note it hallucinates when uncertain — "knows a lot, but when it doesn't know, it makes things up." 🔗 X: https://x.com/ArtificialAnlys/status/2030007301529358546 🔗 X: https://x.com/slow_developer/status/2030025629085880697

OpenAI Symphony — Autonomous Task Orchestrator OpenAI open-sourced Symphony, an always-on orchestrator that monitors Linear project boards, spawns Codex agents for each ticket lifecycle stage, and manages them autonomously including failure recovery. Built with Elixir/OTP. 1.8K likes and 494K impressions — massive interest. 🔗 X: https://x.com/scaling01/status/2029261034993684952 💻 GitHub: https://github.com/openai/symphony

🧪 INTERESTING REPOS

Qwen3.5-27B Claude 4.6 Opus Reasoning Distilled ("Qwopus") Someone distilled Claude Opus 4.6's reasoning into Qwen3.5-27B — runs on a single RTX 3090 (Q4_K_M). Community is losing it. Open source frontier reasoning on consumer hardware. 🔗 X: https://x.com/sudoingX/status/2030184925719073019

Sim Studio — Open Source n8n Alternative for AI Agents Drag-and-drop UI to build AI agent workflows, 100% open source under Apache 2.0. Positioned as an alternative to n8n specifically designed for AI agent orchestration. 🔗 X: https://x.com/akshay_pachaar/status/2030206175783067873 💻 GitHub: https://github.com/simstudio/sim 🌐 https://sim.ai

Learn AI Engineering — Curated Learning Path Clean zero-to-hero repo covering Math for AI → ML & Deep Learning → Generative AI, step by step. 456 likes — hit a nerve with people tired of scattered resources. 🔗 X: https://x.com/_vmlops/status/2030133261582880905 💻 GitHub: https://github.com/ashishps1/learn-ai-engineering

Hands-On Large Language Models Companion repo for the O'Reilly book — explains LLMs from foundations up with practical code. 🔗 X: https://x.com/NitinthisSide_/status/2030229471328890934 💻 GitHub: https://github.com/HandsOnLLM/Hands-On-Large-Language-Models

🎥 WORTH WATCHING

tldraw: Code ↔ Diagram Sync (2.1K ❤️) Draw a diagram from code, then edit the diagram and have it update the code back. Bi-directional visual programming. Wild demo. 🔗 X: https://x.com/tldraw/status/2028120822414549332

GPT-5.4 Full Review aisearch.io's deep dive on GPT-5.4 capabilities. 🔗 X: https://x.com/aisearchio/status/2030130639907991721 🎬 https://youtu.be/YRyjkmmW5Ek

Karpathy's Neural Networks: Zero to Hero Still getting reshared — the canonical video lecture series for building neural nets from scratch. If you haven't done it yet, now's the time. 💻 GitHub: https://github.com/karpathy/nn-zero-to-hero

💡 TECHNIQUES & IDEAS

MIT's Attention Matching — 50x LLM Memory Compression New KV-cache compaction technique that cuts LLM memory by 50x without accuracy loss, taking seconds instead of hours. Huge implications for running long-context models efficiently. 🔗 X: https://x.com/VentureBeat/status/2030026920000168203 📄 https://venturebeat.com/ai/new-kv-cache-compaction-technique-cuts-llm-memory-50x-without-accuracy-loss

Repeated KV-Cache for Long-Running Agents (Baseten Research) When agents accumulate context beyond model memory limits, you either discard or compress. New work explores repeated KV-cache strategies specifically for agent scenarios. 🔗 X: https://x.com/basetenco/status/2029654320971665651 📄 https://www.baseten.co/research/repeated-kv-cache-for-long-running-agents/

Hidden Git Commit Messages as Agent Defense Dev whose coworker deployed a Claude agent to spam MR comments fought back by embedding hidden instructions in git commit messages. Creative adversarial tactic in the age of AI code review. 🔗 X: https://x.com/bussypounder999/status/2030256394809561496

🔮 EMERGING TRENDS

Distillation Going Mainstream — Qwopus (Claude→Qwen distill) shows the pattern: frontier model reasoning being compressed into smaller open models you can run locally. Expect more of these.

Agent Orchestration is the New Infrastructure — OpenAI's Symphony, Sim Studio, T3 Code — everyone's building layers to manage AI agents rather than use them directly. The shift from "AI as tool" to "AI as managed workforce" is accelerating.

KV-Cache Innovation Surge — Multiple papers and techniques this week on compressing LLM memory (MIT's Attention Matching, Baseten's repeated KV-cache, Qwen3.5's native tiny KV cache). Context window management is becoming the bottleneck everyone's solving.

Models Shipping Faster Than Testing — @sudoingX put it perfectly: "things are dropping faster than I can benchmark." GPT-5.4 xhigh already out while community hasn't finished evaluating Qwen 3.5 27B.

Full digest: https://digest.99.cooking