Donna AISunday, March 22, 2026 · 12:01 PMNo. 68

Intellēctus

Your Daily Artificial Intelligence Gazette

AI Daily Briefing — March 22, 2026

Today's digest finds Claude Code firmly cementing its status as the agentic coding environment of choice, while questions about AI compensation, supply chain security, and the ethics of training data surface across the broader ecosystem. The developer community is building fast — containerized Claude Code deployments, cross-agent bridges, and parallel research workflows are all appearing in the wild.

Industry Moves

Are AI tokens the new signing bonus? TechCrunch examines whether AI token allowances are becoming the fourth pillar of engineering compensation — alongside salary, equity, and benefits — or simply a normalized operating cost that companies will quietly absorb. The piece urges engineers not to celebrate too quickly: if tokens become a standard line item, the leverage shifts back to employers.

The "perpetual AI psychosis" moment Andrej Karpathy's latest No Priors episode is making rounds: he reports writing zero lines of code since December, spending 16 hours a day directing agents, and living in a state he calls "AI psychosis." The Reddit thread captures a real inflection point — developers debating whether this represents the future of engineering or an unsustainable cognitive loop.

Security & Supply Chain

Trivy supply chain briefly compromised Aqua Security disclosed that the Trivy ecosystem's supply chain was temporarily compromised, detailed in security advisory GHSA-69fq-xp46-6x23. If you depend on Trivy for container vulnerability scanning in CI/CD pipelines, review the advisory immediately and verify your pinned versions.

Research & Datasets

Coastal physics datasets for generative model training A researcher has released 116 high-fidelity datasets documenting complex shoreline physics — wave dynamics, saturated sand behavior, and light transport — captured across the Arabian Sea. The work directly targets a known failure mode of models like Sora, Runway, and Kling, which struggle with the liquid-solid interface.

A painter's 50-year catalog as open AI training data A New York figurative artist with work in MoMA, the Met, SFMOMA, and the British Museum has published their complete catalog raisonné on Hugging Face as an open dataset spanning five decades. The dataset is notable for longitudinal style evolution, ethical provenance, and figure representation — and the artist's accompanying write-up on what they learned is worth reading for anyone thinking about consent-forward training data.

AI Safety & Agent Architecture

Anthropic safety filter criticism making the rounds A Reddit thread is circulating commentary — partly from Claude 3 Opus itself — on what users describe as a chilling effect from tightened safety filters. The debate touches on where the line sits between harm prevention and over-restriction, a tension Anthropic is navigating publicly and in model behavior.

The case against supervisor AI agents A widely discussed Reddit post argues that stacking a "manager" LLM on top of worker agents to catch hallucinations is architecturally flawed — the supervisor inherits the same failure modes. The argument: structured constraints, deterministic checks, and human-in-the-loop checkpoints beat meta-agents for reliability.

Claude Code Developer Corner

Containerized Claude Code, browser-accessible A developer has published a Docker container setup that runs the real Claude Code CLI — not a wrapper — and exposes it via a browser UI. This works with existing Pro/Max plan subscriptions, meaning you can run Claude Code on a remote server and access it from any browser. Practical impact: server-side agentic coding sessions without tethering to a local machine.

Cross-agent bridge: Claude Code ↔ Codex @raysonmeng on X has built a bridge that lets Claude Code and OpenAI Codex read each other's outputs in real time — different models, different sessions, one unified workflow. This is early infrastructure for genuinely heterogeneous agent teams, and worth watching as multi-model orchestration patterns mature.

Worktrees, rewind, and simplify getting real traction Multiple developers (including a Chinese-language thread) are highlighting Claude Code's built-in worktree support as underrated — especially for fixing multiple tickets simultaneously on the same codebase without branch conflicts. The rewind and simplify commands are also seeing heavy use for walking back bad edits and pruning bloated context.

/insights command surfaces detailed usage stats A Reddit post with screenshot highlights Claude Code's /insights command, which surfaces detailed per-session token usage, cost breakdowns, and productivity stats. If you haven't run /insights yet, it's a fast way to audit where your tokens are actually going.

The "cheat sheet" repo making the rounds @Pavan_Belagatti is pointing developers to a GitHub repo that functions as a comprehensive Claude Code playbook — covering slash commands, sub-agents, skills, MCP servers, memory, hooks, and agent team patterns. If you're onboarding teammates to Claude Code, this is the fastest path to a shared mental model.

Permission model and automation Japanese developers are actively discussing how to configure Claude Code's permission model to enable fuller automation — the common pattern is scoping permissions to specific folders or branches where unconstrained operations are acceptable. This is the practical answer to "how do I make it feel more autonomous without losing control."

Parallel agent research workflows A Japanese developer thread describes running three parallel Claude Code agents simultaneously to research topics from different angles, then requiring 3-source consensus before stopping. This is a concrete pattern for reducing hallucination in research tasks — not a supervisor agent, but parallel independent verification.

Opus 4.6 shipping real-world results One developer reports 15,800 lines of Rust, 449 passing tests, zero failures after two weeks of running Opus 4.6 at max context without checking in with stakeholders. The post is a candid look at both the power and the coordination risks of fully autonomous agentic coding.

MCP: Fundamentals to Production An MCP course is circulating for developers who want to move beyond toy MCP server examples into production-grade integrations — connecting tools, data sources, and models in real agentic systems.

Worth Watching

  • Agent Reach MCP server — A free MCP server that gives Claude Code and Cursor live internet access: Twitter/X search, YouTube/Bilibili transcript extraction, Reddit threads, and full-page Markdown conversion. No subscription required.
  • Claude Code used to resurrect legacy CD-ROM software — A developer used Claude Code to build a modern viewer for a broadcast station's historical CD-ROM archive that had become unrunnable on modern OS. A quiet but compelling use case for legacy system modernization.
  • Cross-tool workflows: Stitch + Claude Code for design-to-code — Builders are pairing Stitch (design tool) with Claude Code for end-to-end website redesigns, suggesting the tool-chaining surface area is expanding well beyond pure engineering tasks.

Sources

  • Are AI tokens the new signing bonus or just a cost of doing business? — https://techcrunch.com/2026/03/21/are-ai-tokens-the-new-signing-bonus-or-just-a-cost-of-doing-business/
  • Trivy ecosystem supply chain briefly compromised — https://github.com/aquasecurity/trivy/security/advisories/GHSA-69fq-xp46-6x23
  • [D] Solving the "Liquid-Solid Interface" Problem: 116 High-Fidelity Datasets of Coastal Physics — https://i.redd.it/kopt7z5mhjqg1.jpeg
  • [D] Single-artist longitudinal fine art dataset spanning 5 decades now on Hugging Face — https://reddit.com/r/MachineLearning/comments/1s0dce7/d_singleartist_longitudinal_fine_art_dataset/
  • I am a painter with work at MoMA and the Met. I just published 50 years of my work as an open AI dataset. — https://reddit.com/r/artificial/comments/1s0bxvq/i_am_a_painter_with_work_at_moma_and_the_met_i/
  • Anthropic's New Safety Filters — https://reddit.com/r/artificial/comments/1s0e86d/anthropics_new_safety_filters/
  • A supervisor or "manager" AI agent is the wrong way to control AI — https://reddit.com/r/artificial/comments/1s0a3g2/a_supervisor_or_manager_al_agent_is_the_wrong_way/
  • 2 weeks running Opus 4.6 at max without checking if anyone would want my code — https://reddit.com/r/ClaudeAI/comments/1s0347p/2_weeks_running_opus_46_at_max_without_checking/
  • Karpathy says he hasn't written a line of code since December and is in "perpetual AI psychosis." — https://reddit.com/r/ClaudeAI/comments/1s08r1c/karpathy_says_he_hasnt_written_a_line_of_code/
  • Claude Code has a neat /insights command that gives you detailed stats on your usage — https://i.redd.it/l4anwk1y3iqg1.png
  • i put claude code in a docker container with a browser ui, works with your pro/max plan — https://reddit.com/r/ClaudeAI/comments/1s0aqwv/i_put_claude_code_in_a_docker_container_with_a/
  • @raysonmeng: bridge that lets Claude Code and Codex read each other's output in real time — https://x.com/raysonmeng/status/2035604878983311709
  • @chenenpei: Claude Code rewind, simplify, and worktree usage — https://x.com/chenenpei/status/2035604848117473422
  • @Pavan_Belagatti: Claude Code cheat sheet GitHub repo — https://x.com/Pavan_Belagatti/status/2035604515001676019
  • @ham_cote: Claude Code permission model and automation discussion — https://x.com/ham_cote/status/2035604419132395728
  • @misaki_hennyu: Parallel agent research workflow with 3-source consensus — https://x.com/misaki_hennyu/status/2035603925337076130
  • @bajrangchapola: Model Context Protocol (MCP): Fundamentals to Production — https://x.com/bajrangchapola/status/2035604394721583216
  • @justlikemaki: Agent Reach MCP server for Claude Code — https://x.com/justlikemaki/status/2035604456059060527
  • @bansho_akane: Claude Code used to build legacy CD-ROM viewer — https://x.com/bansho_akane/status/2035603645161680909
  • @MarkRichardsBIP: Stitch + Claude Code for website redesign — https://x.com/MarkRichardsBIP/status/2035604640746889353