Alex

I'm an Engineering Manager with 25 years in software. I enjoy creating tools that reduce friction for developers. I write about Developer Experience and AI at invoke.dev.

Is it just me, or is Opus 4.8 worse at communicating ideas than previous versions?

Orange is the new purple of AI-generated designs.

One of the more disappointing aspects of agentic engineering's rise is that many industry leaders ignore decades of research on productivity, effective work cultures, and sustainable delivery in favor of unproven tech.

Agentic tools are an excellent solution when short-term output is more important than long-term outcomes.

At $25/$125 per million tokens, Mythos Preview is 5x the price of Opus 4.6 ($5/$25). Anthropic seems to be setting the stage for the next pricing tier that will be closer to $500 - $1,000 per developer per month plus overage. Is the potential 10% productivity gain worth it?

Why does it seem like every new model release starts out feeling like a big improvement, but slowly gets worse over the course of 2 or 3 months to the point they are almost unusable? Are my expectations just changing, or is model degradation a real phenomenon?

I wrote up some thoughts on the impact of AI mandates in software engineering. invoke.dev/ai-mandates/

The silver lining to Anthropic accidentally leaking the source for Claude Code is that actual humans are now reviewing the code, perhaps for the first time.

Is it possible to create an AI coding tool that doesn't suck and instead actually makes you better at writing code?

Why does it always seem like the latest models everyone praises when they are released become dumber and less useful after a few weeks?

The gaslighting from politicians and executives is one thing, but paying for an AI that gaslights rather than following instructions is asking too much.

Once pricing is no longer subsidized and we factor in the costs of process redesign, governance, rework, reliability issues, and skill erosion, I have to wonder if the 10% increase in productivity will justify the total cost of ownership for agentic engineering.

In developer experience, we used to talk a lot about the importance of maximizing Flow or deep work time for engineers. This was one of the core productivity metrics we measured. Is Flow even possible for the human in agentic engineering, and does it still matter?

RIP Sora - You won't be missed.

What is likely to happen when most Fortune 500 companies and software engineering organizations are completely dependent on AI to run their business, and AI subsidies have run out?

@getdx.bsky.social tracked 40 companies over 15 months to understand real world impact of AI. Usage is up 65%, but despite the hype, productivity gains are realistically closer to 10%, not 10x. newsletter.getdx.com/p/ai-product...

@addyosmani.bsky.social makes a good case for being deliberate with AGENTS.md. Auto-generated context duplicates what the agent already discovers on its own, and research shows that hurts performance. Save it for tooling gotchas and non-obvious conventions. addyosmani.com/blog/agents-md/

Can we skip the "AI-native" phrasing? Next thing you know, companies will be putting it in job postings. It is a nonsense term.

Don't anthropomorphize computers. They hate that.

Hey @mubashariqbal.com please add: go.bsky.app/32NPHXw

Hey @blueskystarterpack.com please add: go.bsky.app/32NPHXw

Agentic coding tools need project knowledge, conventions, and domain expertise loaded into every session. I've settled on a four-layer context architecture. One source of truth that is portable across agentic coding tools. invoke.dev/context-engi...

Doubling down on asccli.sh for indie devs. What should I add or improve next?

This article is lacking context. These cuts come on the heels of articles in both Inc and Wired about poor employee morale at Block and frustration with Jack. www.inc.com/kevin-haynes...

Oh no... what have we done...

Tech CEOs love framing AI as an arms race. Someone should remind them that the last arms race ended with a superpower going bankrupt.

Automation often shifts toil rather than eliminating it. Time saved writing code reappears in review, rework, and bug fixing. AI is a tool, not an alternative to humans. Position AI as a quality-of-life improvement, not a productivity multiplier. The honest framing actually gets better adoption.

If you've tried AI code review and turned it off because the comments were useless, you're not alone. I documented the tactics I found most effective for improving the signal-to-noise ratio in GitHub Copilot AI-assisted reviews. read.invoke.dev/p/ai-assiste...

Teams keep reaching for AI to generate more code, even though the actual bottleneck is review. Generating code faster in an already constrained review process doesn't improve throughput. It degrades it. Fix the gate before you widen the pipe.

I've had enough AI "woo-woo." If this technology is going to be meaningful and useful we can't surround those discussions with a fog of mysticism. biblioracle.substack.com/p/can-we-ple...

I spent some time coding without any AI this evening. It was very refreshing.

I'm pretty sure I know. They aren't. We're not going to really be able to make a lot of progress in dealing with the implications of this tech unless and until we get rid of all this "woo-woo" talk about LLMs. Anthropic pushing this line is PR, unserious.

Vibe coding is the alchemy of software engineering — mixing substances and hoping for gold, with no understanding of why anything worked and no guarantee that the success can be replicated.

When critical systems inevitably start breaking due to LLM slop, who do you think companies will look to hire? More vibe-coders? Or engineers who actually understand the technology and can produce high-quality software? Vibe-coding clearly can't even deliver *two* nines of reliability 🫢

AI tools are like slot machines: unpredictable payoffs that produce enough dopamine to keep you pulling the lever. It can be addictive, and the rewards are hollow. Invest in your ingenuity and critical thinking. Use the tools responsibly.

Vibe coding is to software engineering what DoorDash is to cooking.

Anthropic's research (www.anthropic.com/research/AI-...) found that developers using AI scored lower on comprehension tests. AI tools require critical thinking to use effectively, yet they erode that very skill. I wrote about what this means for the adoption of AI coding tools.