Matt Makai | Full Stack Python | Plushcap (@fullstackpython) header
Matt Makai | Full Stack Python | Plushcap (@fullstackpython) profile picture

Software dev. Expertise in DevRel & DX. Currently building https://www.teapawt.org/t.co/qpRK3z8M5s. Prev DevRel leader @twilio @digitalocean @AssemblyAI @launchdarkly

Virginia, USA
Joined January 2015
Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)Photo by Matt Makai | Full Stack Python | Plushcap (@fullstackpython)
Matt Makai | Full Stack Python | Plushcap retweeted
had to do it, meme was too good, periodic table live now in @GoogleAIStudio : )
123
95
24
2,071
60,099
Matt Makai | Full Stack Python | Plushcap retweeted
most developers would do much better just using one AI chat a time and actually think about what they are doing too many of y'all larp
408
367
105
7,853
277,303
Replying to @craigsdennis
I did not, until now! Also got the D&D one on there
1
4
Replying to @craigsdennis
this is wild Craig! I was just playing that Transformers pinball machine a week ago because our local arcade has it. I don’t know if I want to use AI to become better at it or just keep pretending that I am improving by trying it over and over 😅
1
1
96
Hmm, so @pidotdev is the future one harness for every fat model? 🤔
Fascinating article that shows that LLMs eat scaffolding for breakfast. It's gonna be fat model and thin harness and context is what will matter most. "You cannot turn off thinking on Opus 5.5, and Fable 5.1, and recent models increasingly narrate their progress and check their work as they go, which makes some harness elements redundant. Each new model lets you move more of the decision-making out of the harness and into the model itself. A simpler loop is also easier to move to the next model; in migrations we have helped customers with, rigid multi-agent structures were often what stopped the new model from doing more of the work" claude.com/resources/article…
1
682
Replying to @ceoonlocalhost
yes exactly, completely agree with this framing. I think we're there already with existing AI coding tools for software of a certain size (but not for companies with dozens or thousands of applications/microservices interacting)
1
1
11
Replying to @ceoonlocalhost
Usage flattened because there was only so much I could do in a day after active building on plushcap.com, even when running agents to optimize SEO, background tasks, etc I don’t know if it got “better” since thats subjective but Plushcap is closer to my vision for what it can be. Good software isn’t about infinite growth its about taste, vision, and constant improve based on user feedback. But even those have theoretical token limits!
1
1
21
Gem of a thread on token usage. Based on 147 heavy Codex users but matches my own development patterns with a heavy ramp-up followed by typical S curve plateau as there's only so many tokens I need per day even if it's way higher than previous non-agentic coding token usage levels.
OpenAI just reported $10B in quarterly revenue, up 50% in three months. Claude and I pulled a year of daily token counts from 147 public Codex profiles to see it up close. The surprise: usage soars as people adopt agents, then levels off fast, well below plan limits. 🧵
1
1
7
1,392
Man platform engineering seems to be having a moment recently. On target for ~150% growth this year in new features, content, and products across DevTools companies. My hunch is that the rush to adopt as many AI coding tools as possible in early 2026 led to governance challenges for many eng orgs. Also probably just a natural progression from "figure out all at all costs" to "ok yea those are real costs we need to streamline these processes and tools" @GergelyOrosz's had a bunch of solid qualitative insight on platform engineering and related subtrends (AI governance, internal dev platforms, source control) in The Pragmatic Engineer newsletter as well (and the Platform Engineering Handbook) Data: plushcap.com/trends/monthly/…
5
2
11
1,210
Matt Makai | Full Stack Python | Plushcap retweeted
“Now that we can build everything, maybe we should not?” A great talk from @mitsuhiko, CTO of Earendil, who instead of software factories suggests to build software ateliers. Or sort of workshops. Less software, but better, more ambitious software. How many projects have you started recently? Exactly. Over the past year, Armin said he started 141 projects. But how many does he regularly return to and maintain? Just 10–15. And that number hasn’t really changed. Meanwhile, his agents went from running for a couple of minutes to 40+ minutes on longer tasks. The models got dramatically better, while the core tools in his Pi coding agent barely changed. I think that adds to the high-functioning anxiety that I’ve written about recently. Because though we can build much more, our attention hasn’t scaled with the models. What to do? Armin makes a few good points: - Learn when to push back against agents. - Don’t confuse a completed task with good software. - And don’t abandon collaboration and shared standards just because it’s now easier to build your own version of everything. I think there’s also a bigger question here: if the cost of building keeps falling, how do we get better at deciding what deserves our attention? Maybe that’s where the real skill is moving. From being able to build something to knowing why it’s worth building, and caring enough to keep it alive.
6
10
43
7,933
Agent evals weekly trend data in Plushcap was +8% week-over-week but it's a noisy trend, basically flat the last few weeks with just some ups and downs along the way: plushcap.com/trends/weekly/a… This is a trend that will matter eventually when enough companies have agents in prod.
1
434
Agent evaluations as a developer trend ticked up slightly (+8%) across companies last week and this post by @Cometml is one of the best recent eng blog posts on agent evals. They created an incident handling test environment to evaluate defensive remediation agents: comet.com/site/blog/optimize… Their team set up ten incidents (with 4x variants each) in a Docker-based payments stack w/ Prometheus, Loki and @grafana supplying metrics, logs and alerts. Then they deliberately broke the stack to trace how agents investigate and responded. Definitely a worthwhile read.
9
1
17
1,494
Replying to @clairevo @chatprd
Ughh I’m so sorry you have to deal with this stuff Claire. Hang in there and keep on showing how to build a great product, even after these new constraints have to be put in place due to bad actors 🫡
1
1
396
nice! the 19% faster with 20 tool calls is interesting. does that scale linearly over SSE in your testing, so 5 tool calls ~5% faster, 50 tool calls ~50% faster? or there’s an inflection point past a certain number of tool calls?
121
Matt Makai | Full Stack Python | Plushcap retweeted
Friends don't let friends use slop WebSocket client code in their apps. Check out the new official WebSocket helpers in the @OpenAIDevs SDKs - Responses, GPT-Live, and Realtime APIs all supported, efficient, and properly typed. Also major latency wins on tool calls 🧵
2
4
15
906
Matt Makai | Full Stack Python | Plushcap retweeted
By the way, in case anyone was still wondering whether they are completely useless, Binance holds SOC 2 Type 2, ISO 27001, and several other certifications.
161
456
90
4,795
254,742