I'm back! If you've noticed a large gap between my last reflection and this one, that's because I have just returned from vacation. I had the incredible opportunity to travel Austria and Germany with my girlfriend, and then visit the UK for a family Indian wedding on my mom's side.
Returning home
Being back from vacation, I needed a vacation from my vacation. So my first few days were lightly scattered with work but not the biggest things. I signed up for webinars in PlatformCon before leaving, which is a convention for platform engineers. I thought it'd be interesting, so chose a few that I thought would be useful or more accessible to me with no background in the space.
Platform engineering webinars
The first webinar was from a company with LOTS of data on how AI is being used in workplaces. They have data from 1300 companies, 275k devs, 90M PRs. All from AI coding tools, agents, IDEs, and more. The data was very interesting because the amount of PRs being made are up drastically yet the epics completed are not up nearly as much. Epics are an interesting metric for productivity here, but I do see how it's hard to measure the "real" impact of AI without something like an epic that is supposed to measure a tangible business outcome or goal. Interesting webinar and it reinforced the idea of context being so valuable to these agents, since it equaled more "useful" code being shipped. Although they weren't clear on how impactful that code is, but I think in general the quality of the code will be better with context. Especially when shared logic is required over a whole ecosystem of code.
Honestly, the rest of the webinars were either less interesting or got postponed due to technical issues, so I'll skip those in my reflection.
Planning to meet cool people
Other than that, I've been on the grind for meeting with people who I've wanted to chat with. I have someone from Scotiabank that I got connected to through an old connection, someone from a conference, and a small business owner who I've been talking to for a few years now. Once I have those settled, I want to reach out to a startup founder that I've loosely kept in touch with to see how they're doing. It's always interesting to get perspectives from across the industry and also just keep in touch with people I like talking to.
An early user for startups
In addition to my own outreach, I joined Boardy Pro's founding group and received access free for life. It's a startup that aims to be an AI Superconnector, helping connect people for networking, sales, and more. Their pro version doesn't just make the intro but helps it achieve its goal, so I think sales is the clearest use case because it'll stay on the thread and push things if they stall. As part of this, I also opted to join their Advisory Board which is for pro users to give feedback. So far, I'm letting it do its thing by connecting me with people in Toronto who can help with my new grad job search. Banking, fintech, startups, honestly I'm just interested in technology as a whole so it doesn't really matter where I end up. Those are just the places I'm more familiar with, but I'll get Boardy to expand its outreach for me in the coming month. I'm very curious to see how it performs since it hasn't made any connections yet but says it's working on a couple for me.
Organizing... everything
Finally, I've been really trying to get my crap organized. For example, I already made a sprawling obsidian folder to keep track of everything personal, including these blog writeups. I had a habit of leaving emails unread as a sort of "todo" or "look at later" flag, but just moved them all into obsidian. It's git-tracked and private so I like having it as a clear place for everything I deem important to be stored. The graph on obsidian helps me traverse what I need if I'm curious (although just going through the folder structure so far has always done the trick), and I'm able to organize things very nicely so I have a clear separation of todo, to remember, to come back to later, and all the other things that I want stored and easily searchable. Some of the emails that I moved were cool tech articles that I might want to revisit, goodbye messages from Scotia that I saved because they show the positive impact that I made, and some opinions from my mentor there that I deemed important for me to remember as a developer.
Honestly, pretty good week back in terms of being productive but also letting myself rest. I think with new grad applications coming up, I'm realizing that keeping in touch with my network is even more important, as well as grinding Leetcode so I'm not dropped from round one interviews. I've got some cool personal projects that I've been exploring too and I'll hopefully write about them next week. Cheerio :)
Webinar notes (highly unstructured)
After watching these webinars, I unfortunately don't have any interest in summarizing them for my blog. But in case they're interesting or helpful to anyone else, read about them below!
20 million PRs later: The truth about AI productivity
- start by talking about the real ROI and why we aren't seeing it yet? confusing cuz i think we're def seeing some at least in personal use
- they have data from 1300 companies, 275k devs, 90M PRs, from ai coding tgools, agents, IDEs,
- they're seeing a steady increase where now the median company has 70% of users using AI
- looking at autonomous agents where you give it a goal and it makes PRs, commits, etc. on your gbehalf without you even entering the coding scape. the agent does all the work and the person then just reviews the PR. median company had a massive spike from de 2025 ardound 2% to 8.7% now. last year, most were 0%. the top companies at 90th percentile arehaving 34.8% now using autonomous agents for PRs
- high adoption is clearly driving PR throughouput, from 0.96 per week per engineer to 2.43 now
- even if PRs are going up, they see less gains for how many epics are completed (I8 -> 10.2 epics / 100 engineers / week)
- interesting metric bceause some compnies don't do epics, some use difefrent naming (do they accoubnt for that?)
- what if epic scope is changing with AI use? does the scope simply go up because they know they can do more?
- they then talk about why the gap might exist
- bottlenecks are agent autonomy, token economics, and quality concerns
- human attention is hard, they see a massive bottleneck when managing 3 or more agents at once. manaigng 1-2 agents, is most common. that said, this is them looking at what people do. there is a chance that people would be best doing more or less but we just can't see it, since this is a behabvioural thing and not a "people's attenbtionb fall off with 3 ore more", but rather "most people don't bother with 3 or more because they likely think they can't handle it"
- costs are accelerating where 90th percentile hasd 301M toekns per user per week. 50th or median percentil only has 35M.
- they notice that the cost of tokens grows faster than hte productivity gains, looking at deciles where PRs grows slowly and token costs seem exponential
- averaeg dev is about $150/month while p90 dev is more like $1000/monthj so costs are massively increasing at the top end. important to manage the costs across the org and ensure that the costs are truly bringinbg benefits into the org with outcomes
- they don't currently see any quality issues at scale because of AI agent use, purely based on the past year. the amount of PRs that were reverted because of a quality issue have basicallyt not change at all
- they're seeing that agent-made PRs are merged at lower rates than human PRs, by a good 10-20%.
- to improve ROI
- adding more context file investment equals more code shipped -> teach your agents more
- measure for your outcomes and not just activity
My Click. Run. Validate. How to add self-service test execution to Backstage
- platforms help internal developer teams move faster and abstract away complexity from them to streamline their work
- most platform teams work on an internal dev toolbox
- quality and security are hard, main challenges are available for everyone who needs access, standardized and scalable
- testkube is an open testing platform that uses AI to understand your tests, environments, history, and dependencies
Making observability a first-class platform concern
This workshop goes over how to design, evaluate, and operate OpenTelemetry in production-oriented kubernetes environments. It wasn't able to be completed due to them having technical difficulties, so they rescheduled it. I tried it myself, and figured out a couple things in the beginning. After that, I realized that I am sorely lacking the background for the rest. When it asked for "fields that are missing", I tried cross-referencing with other config files, but the fields missing were ones that I simply didn't know exist.