🦸 Community Superhero

Our community superhero this week is Brianna Heffernen, engineering manager on the security side at Datadog. Brianna has a sharp focus on both security and developer experience, which is one of the hardest places to operate because you’re asking people to slow down for reasons they can’t see (we’ve all been on the receiving side of this). She just published a great article on how Datadog enforces software supply chain policies, and I’m here for it. It’s exceptionally well written and comes from an obvious background in tough engineering. Thank you for everything you do, Brianna!

💯 Spotlight

If I asked you to write me a Lambda function, you might ask me “what kind?” A couple years ago, I would have looked at you like you were crazy. But as it turns out, that’s a totally valid question! Ran Isenberg published a piece describing what AWS Lambda actually is in 2026 and talks about the five distinct types of functions. Classic functions, durable functions, managed instances, microVMs, and now web functions. It’s wild to see how far we’ve come and how we’re getting the right compute options for whatever we’re building.

🔥 My Favorite Content

If you run anything agentic in production, go read Ryan Skidmore’s write-up on how Cloudflare cut their inference bill by 68%. Their AI code reviewer had crept from $1.19 per review to $2.43 in only 4 months with no architectural change, and they got it down to $0.77. They were also able to cut review time from four minutes down to a minute and a half 🤯. Surprisingly, the biggest optimization was moving a prompt caching breakpoint so they could cache across repos. As with most Cloudflare articles, this is full of engineering goodness that makes you want to sit down and get to work optimizing your own code. Great work!

Dave Hall wrote a wonderful blog on the newly released EventBridge Custom Buses last week. I loved his statement “give every team its own event bus and you end up with a mess, not a mesh.” 🤣 Dave walks through the new bus vs the “classic,” and it’s surprising how many new features v2 has: FIFO per event group, built-in retention up to 365 days, content-based dedup, and JSONata transforms on the way in. Dave goes on to talk about the costs, which is different than what we’ve seen before. Custom buses charge per GB in and per GB out per subscriber, meaning if you’re publishing 1MB payloads you could be looking at roughly 12x the cost. Dave also points out that retries dropped from 185 over 24 hours to 5 over 5 minutes, which is going to surprise quite a few of us who make the switch.

This next one has a vendor pitch in it, but is also has meaningful content around building AI capabilities at scale. Madeleine Quinn over at Ably talked about backpressure in AI streaming in a post that covers how you should change how often tokens move under load. She proposes two tests: throttle one client in dev tools and check its lag stays bounded, then reload mid-answer and see whether you get the whole response so far. It’s a great perspective I can’t claim to have thought about, but clearly she does and she makes great points.

A little more philosophical, Brian Tarbox argues that we need fair witnesses again. I’ve never heard the term before, but Brian does a great job explaining what they are and makes a solid case for bringing them back… especially in the age of AI. Brian points out that since anyone can generate convincing video and voice, trust has to shift from who a source is to how it knows. He ties that into agent development and how memory can quickly become a game of telephone if you don’t keep the original memory immutable. This was a really cool read, and a great perspective on truth telling.

💡 Tip of the Week

I love this tip from Christoph Werner about catching issues early with Claude Code by enabling the “You should know” plugin. I saw this last week and started working with it and I love it. Since he didn’t say the command to enable it, here it is: /plugin enable cc-plugin-you-should-know@builtin

Last Words

A big motif through this issue is knowing what you’re paying for. Five Lambda modes that bill five different ways. A cache breakpoint worth sixty cents a review. A per-GB price that flips at 4KB. An agent memory that edits itself a little every time you read it. The tooling keeps getting more capable but our understanding is slowly dwindling. Those who are doing particularly well are the ones that are actively maintaining their processes, not living in a “one and done” mentality.

That’s my take on the week, but what’s yours?

What did I miss? What made you nod along (or 🙄)? Hit reply if you’re reading the email. Prefer socials? Ping me on Twitter, LinkedIn, or email.

Happy coding!

Allen