Welcome to 2026! This is my 42rd monthly newsletter, which means I've been writing this for 3.5 years now (here's the full archive).
Over the past year, AI coding agents have completely reshaped how we work — from vibe coding, to agentic coding with human review, all the way to fully automated multi-agent orchestration. We're all at different stages of this evolution, and I put this edition together to help you navigate it. 👇
Andrej Karpathy on the State of AI
Andrej Karpathy has been one of the most insightful voices in AI. I featured his 2025 LLM Year in Review in my previous newsletter, and since then he's shared two more posts that are well worth your attention.
📝 "I've never felt this much behind as a programmer"
Earlier this month, Karpathy posted on X that he'd "never felt this much behind as a programmer." The post went viral and resonated deeply across the developer community.
He describes a new "programmable layer of abstraction" involving agents, subagents, prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, and more. In his words:
"Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession."
If even Karpathy feels behind, the rest of us can feel a bit less guilty about struggling to keep up 😅
For a counterpoint, Maximilian Schwarzmüller recorded a reaction video with a reassuring "Relax" message. I think the truth is somewhere in the middle — things are moving fast, but there's no need to panic 🙂
📝 Random notes from Claude coding
Most recently, Karpathy shared a detailed thread with practical observations from weeks of heavy Claude Code usage. This is the most actionable of the two, and I found myself nodding along to many of his points.
Here are the highlights:
- Workflow shift: he went from 80% manual coding to 80% agent coding in just a few weeks. "I really am mostly programming in English now."
- Agent pitfalls: models make wrong assumptions, don't seek clarifications, don't push back when they should, and are "still a little too sycophantic." They'll implement 1,000 lines of bloated code, and when challenged, immediately cut it to 100.
- Agent swarm hype: "too much for right now" — watch models "like a hawk" if you care about your code.
- Tenacity: agents never get tired or demoralized. Watching one struggle for 30 minutes and come out victorious is a "feel the AGI" moment.
- Speedup vs. expansion: the main effect isn't just speed — it's doing things that "wouldn't have been worth coding before."
- Key tip: "Don't tell it what to do, give it success criteria and watch it go." Shift from imperative to declarative.
- Fun factor: programming feels more fun — drudgery removed, creative part remains, less blocked/stuck.
- Atrophy warning: he's already noticing a decline in his ability to write code manually.
His TLDR: "LLM agent capabilities have crossed some kind of threshold of coherence around December 2025 and caused a phase shift in software engineering."
AI Workflows and Tools
So, with the big picture in mind, let's look at what builders are actually doing with these tools.
📝 How the Creator of Claude Code Uses Claude Code
Boris Cherny, the creator of Claude Code, shared his workflow in a viral thread that was covered by VentureBeat, InfoQ, and Fortune.
Here are the key takeaways from his workflow:
- Runs 5 parallel Claude sessions locally, plus 5-10 on claude.ai
- Each session uses its own git checkout (not branches or worktrees)
- Starts in Plan Mode, iterates until the plan is good, then switches to auto-accept mode
- Maintains a CLAUDE.md file per team to document mistakes and best practices (~2.5K tokens)
- Uses PostToolUse hooks for auto-formatting
- Most important tip: "Give Claude a way to verify its work" — tests, browser, simulators
No exotic customization. No clever hacks. He also accepts that 10-20% of sessions simply get abandoned.
For more details, there's even a dedicated site: howborisusesclaudecode.com. Definitely worth a read if you use Claude Code.
📝 Shipping at Inference-Speed
Peter Steinberger, the prominent iOS developer and PSPDFKit founder, takes things even further. His claim: he no longer reads the code his AI agents generate.
Instead, he manages 3-8 simultaneous projects, commits directly to main, and relies on AI agents to iterate and validate. Software development, he argues, is no longer limited by coding ability — but by "inference time and hard thinking."
I always find it fascinating when very experienced devs like Peter claim they no longer read or review the AI generated code on their solo-projects. Yet, many others still advocate that human-in-the-loop is necessary. Who's right? 🤔
This tension — between "ship without reading" and "human validation is essential" — is one of the most interesting debates heading into this year.
AI Agent Orchestration
Beyond single-agent workflows, some developers are pushing into multi-agent territory. Here are two approaches at opposite ends of the spectrum — and both are fascinating.
📝 Gas Town: Multi-Agent Orchestration
Steve Yegge's Gas Town is a Go-based orchestration system that lets you coordinate 20-30 parallel Claude Code agents using tmux. It features 7 distinct worker roles and runs on a git-based work tracking system called Beads.
What I find most useful is Yegge's 8-stage maturity model for AI-assisted coding:
- Stages 1-2: No or minimal AI (autocomplete, sidebar chat)
- Stages 3-5: Single agent with increasing trust and automation
- Stages 6-7: CLI, multi-agent, hand-managed (3-10+ parallel instances)
- Stage 8: Building your own orchestrator
Most of us (myself included) are somewhere between stages 3-6. Gas Town is for stages 7-8. Fair warning: running heavy sessions can cost $100-200/hour in API fees 😅
📝 The Ralph Wiggum Technique: Simple Agent Loops
At the other end of the spectrum, Geoffrey Huntley's Ralph technique is brilliantly simple — just a bash loop that repeatedly feeds Claude a prompt file:
while :; do cat PROMPT.md | claude-code ; done
Each iteration gets fresh context, and memory persists via git history and progress files. The results can be remarkable: a team at a Y Combinator hackathon produced 1,100+ commits across 6 repos overnight for ~$800 in AI costs 🤯
Anthropic later built an official ralph-wiggum plugin, though Matt Pocock argued that the plugin misses the point: a proper Ralph loop gives bash control over the agent, while the plugin inverts this (letting the agent control the loop, leading to context rot). If you're curious, this article argues why the original bash loop might be preferred.
Anthropic's ToS Crackdown
Earlier this month, Anthropic deployed safeguards that blocked Claude Pro/Max subscription tokens from working outside the official Claude Code CLI. Overnight, third-party tools like OpenCode stopped working — with no warning.
Anthropic's rationale: third-party tools had been spoofing the Claude Code client identity, generating unusual traffic without telemetry, and making debugging and support impossible. This violates their Terms of Service.
There's also an economic angle: the $200/month Max plan provides unlimited tokens through Claude Code, while the same usage via API would cost $1,000+. Third-party tools removed the artificial speed limits, enabling overnight autonomous loops.
The community reaction was strong — subscription cancellations, front-page Hacker News discussion, and criticism from prominent developers, including this entertaining take from Primeagen:
Regardless of where you stand, the takeaway is clear: if you rely on Claude's subscription plans, be aware they're scoped to first-party tools only.
This is a good reminder that we should never be too reliant on a single tool or provider. Having alternatives is always a good idea.
OpenCode: Open-Source Alternative
Speaking of alternatives, if the Anthropic crackdown has you looking around, OpenCode is worth knowing about.
OpenCode is an open-source AI coding agent with a terminal UI that supports 75+ LLM providers — Claude, GPT, Gemini, local models, you name it. It's been growing fast: 56K+ GitHub stars and 450+ contributors.
Key features:
- Provider flexibility: swap providers or bring your own API keys — no lock-in
- Privacy-first: no code storage, suitable for regulated environments
- Client/server architecture: enables remote sessions (Docker containers, mobile control)
- LSP support: language-aware editing, multi-session, shareable links
Having briefly tried OpenCode myself, I quite like it, especially as it already supports useful features such as custom slash commands and skills, meaning I can reuse many of the workflows I already built for Claude Code.
Check it out here:
- OpenCode (source on GitHub)
- OpenCode vs. Claude Code comparison
Claude Code Resources
If you're using (or getting started with) Claude Code, here are some of the best resources I came across this month:
- 📝 The Complete Claude Code Tutorial — a viral X thread (4.7M views) by Eyad Khrais. Core message: think before typing. Plan mode outperforms ad-hoc prompting 10 out of 10 times.
- 📝 A Guide to Claude Code 2.0 — a deep technical guide covering Opus 4.5 workflows, sub-agents, MCP servers, and hooks. Interesting note: the author finds GPT-5.2-Codex superior for code review, while Claude excels at code generation.
- 📝 Claude Code in Action — free 21-lesson course from Anthropic (also on Coursera). Covers everything from basics to Hooks and the SDK.
- 📝 Todos → Tasks in Claude Code 2.1 — the latest Claude Code update introduces session-scoped Tasks (replacing Todos) for complex dependency management and parallel sub-agent coordination.
Latest from Code with Andrea
📝 My 2025 in Review: Freefall and a New Direction
At the start of this month, I published my 2025 retrospective. It was an honest look at a challenging year: reduced content output, declining traffic and revenue, and the broader headwinds hitting coding educators everywhere.
But the article is ultimately forward-looking. I'm pivoting Code with Andrea towards agentic AI coding — using tools like Claude Code to build Flutter apps faster and smarter. I won't be teaching traditional Dart/Flutter tutorials anymore, but I'll continue to cover Flutter in this newsletter and focus my content on AI-assisted development workflows.
If you haven't read it yet, I'd love to hear your thoughts:
Until Next Time
This is by far the most AI-heavy newsletter I've ever written. AI coding agents have crossed a threshold, and I find that learning this new abstraction layer is well worth the effort.
My advice? Start small. Try Claude Code or OpenCode on a side project. Give the agent success criteria instead of step-by-step instructions. And review everything it produces — at least for now 🙂
I'd love to hear where you are on the AI coding journey! Let me know on X (Twitter), LinkedIn or BlueSky.
Thanks for reading, and happy coding! 🎉
