Instagram story viewer> @jackrudenko> Posts
Jack Rudenko(@jackrudenko)'s profile avatar HD Profile

Jack Rudenko

@jackrudenko

139
followers
184
following
👨‍💻 CTO @madappgang AI agents are evolving SaaS. Claude Code enthusiast sharing dev tips&trends. MedTech, IoT, high-load | AWS 🤷🏻‍♂️ | go | dingo
POSTS STORIES REELS TAGGED
Every company I know is trying to build a new AI engineering team right now, and all of them are struggling, including mine. Each one has a few AI champions who look superpowered. The rest of the team is where it falls apart. At 10X Labs, we're 50 engineers on Claude Code, and we're somewhere in the middle of that path. Some things worked. More things didn't. Our first plan was to let the champions lead and let everyone build their own setup. A few months later, every laptop had a different agent, and "the frontend skill" meant something different to each engineer. Our AI champions are great. They're also the reason we had 50 different agents. That's one story among many: AWS, Spotify, Replit, LinkedIn, and even the Claude Code team managed by Boris. We've been collecting the rest from our own teams, from partners, and from other companies trying the same thing. At AI Engineer Sydney, I'll share where we are on the path, what we know so far, and what we still don't. I'll also release the tool we built for the config part as open source on stage. 7 to 8 December, Hilton Sydney. If you're the person who has to make AI work for a whole team, I'd love to have you there. Early bird pricing ends on 9 October. Please use my discount code to get $200 off tickets and 100$ off streaming: jackaieng26syd #cludecode #AI #magus by @jackrudenko
3
a day ago
Download
Opus 5.5 is amazing, but what I found is that the smarter the model is, the less likely it is to follow your instructions. That is a scary progression. #opus5.5 #ai #claudecode by @jackrudenko
0
2 days ago
Download
Three weeks ago, I wrote that Fable gives us the same diff as Opus and costs about five times more. This week Anthropic made that post wrong, and I'm happy about it. The new Opus and Fable prices landed, so I opened our model catalog to copy them into MadBench. Screenshot attached; the annotations are mine. I expected the usual small change and a slightly bigger invoice. The input and output rows did move a little, which is roughly what I expected. What I didn't expect was the cached read row, which dropped on both models, and on Fable it dropped hard. "Why do you care about the cache row, Jack?", you'll ask me. Because that row is most of what a coding agent costs us. In a chat, you type a question and the model reads it once. An agent in Magus and other agents works differently. On every turn, it sends the whole conversation again, the system prompt, the CLAUDE.md, every file it opened, and every tool result so far. The model reads the same files over and over, and a cached read is the price of each reread. Like a contractor who rereads the whole brief every morning and bills you for the hour. On our invoices at MadAppGang, the cache already saves 20 to 40%, and in a long Magus session most of the tokens are rereads. So the five times in my Fable post came from the old cache price. I don't know the new gap yet. We're running the same tasks through MadBench again this week, and I'll post both invoices. I have no idea how they did it. My read is that a drop like this takes months of inference work, and this release it got the same effort as making the model smarter. That deserves a thank you, so here it is. #AI #opus5.5 #claudecode by @jackrudenko
0
3 days ago
Download
This week I couldn't scroll for ten minutes without an AI influencer calling Jev a revolution. I watched the demos. They showed a problem we solved in 2018. Only now it runs slower and costs more. So what is Jev, really? You give it a piece of text and a question with fixed answers, and it picks one. Is this email urgent, yes or no. Which team gets this ticket: billing or technical? It answers in under half a second. The cost is a fraction of a cent. That's faster and cheaper than ChatGPT or Claude. But that comparison hides the real story. Sorting text into buckets is called classification. In 2018 Google released BERT, a model that reads a sentence and decides which bucket it belongs in. A model like that has been hiding spam from your Gmail inbox for years. BERT is a language model. But it isn't a large one. It's small enough to run on a laptop. Once trained on your data, it answers in about a tenth of a second on a normal CPU. No GPU. No cloud. Jev runs only on TypeSafe's servers, behind a waitlist. Its speed is measured against chat models that were never built to sort tickets. Asking Claude to return the word "billing" is like hiring a lawyer to read a room number off a door. Most of the demos showed Jev solving the same problem BERT was built for. Parallel, a company that runs classifiers at scale, tested this last week. Its own trained models beat Jev on two of three tasks, and at a lower cost. Jev playing Pong at 227 milliseconds per move looks impressive next to Claude Haiku at 2.5 seconds. Then you watch the video below. That’s a small model, not a language model, playing Pac-Man on my slow laptop with no cloud involved. The project I created a couple of years ago. It still has real uses. I described it in the article. Link in the first comment. #AI #Jev by @jackrudenko
7
8 days ago
Download
I’ve got the Threads notification bell turned on for exactly one person, and it is not a client or a boss. It is Boris Cherny. I read everything he posts about Claude Code the way other people follow football. Half of it is jealousy. The man has unlimited tokens. I have a billing alert. Fine, the stalking part is a joke. The next part is not. He wrote that he keeps effort at xhigh for most work, and max for the hard tasks. It is a simple sentence, but it changed how I set up my agents. For a year, I kept picking models. Sonnet here, Opus there, and when the answer was bad, I would just go one model up and pay more. Effort gives you a second question to ask. If the agent had everything it needed and still got the logic wrong, that is the model. If it skipped a file, never ran the tests, or stopped halfway through a refactor, that is effort. Same bad result, different reason. I had been paying for intelligence when what I needed was patience. So now I set two things instead of one: which model, and how hard it works. Then there are subagents. My session runs high, so the small agent that greps a repo and reports back runs high too. It does not get smarter from that. It just costs more to read filenames. We fixed Claude Code by adding this in Claudish, where every agent carries its own effort level. I hope Claude Code does the same, and I do not think anyone will argue with that. Boris, if you are reading this, I am not following you. I am benchmarking. #AI #ClaudeCode #Fable by @jackrudenko
2
9 days ago
Download
Buying everyone a Claude subscription does not give you an AI engineering team. We tried. You get one engineer flying and a lot of very expensive autocomplete. The gap between your best AI engineer and the rest of the team is the real problem. The top teams are fighting it right now too. AWS, Spotify, Replit, OpenAI, and the Claude Code team are all building their AI engineering teams as we speak, and they struggle like everyone else. We learned a lot from them, and from our own mistakes. I'm putting their lessons next to ours at AI Engineer Sydney. What worked, what failed, and where all of us are still guessing. 7 to 8 December, Hilton Sydney. Code jackaieng26syd gives you $200 off conference and leadership tickets, and $100 off streaming. Send this to whoever signs your AI invoices. #AI #ClaudeCode #AIEngineering #TechTok #Sydney by @jackrudenko
8
9 days ago
Download
A developer on the team installs the skill you wrote, follows the instructions, and the agent never uses it. 400 skills on that machine. I assumed ours had been cut from the listing. So we captured the listing the model actually received and read it. The skill was there. The name was there. Only the description was gone. That is what happens over the character budget. Claude Code removes descriptions, whole, and keeps every name. Nothing in the listing marks which lines lost one. In our setup, 415 skills installed and 332 of them arriving as a name. Which changes the problem. You are not trying to get your skill back into the listing. You are trying to make one word carry it. So we measured that. wf-7742, a bare line in a 415 skill listing, was used in 1 run out of 10. Renamed to reconcile-ledger, nothing else changed, 10 out of 10. And when a skill must not be missed at all, take discovery out of it. We gave a subagent the skill in its config and removed its Skill tool completely. No listing reached it. The skill body arrived in its first message at startup. 5 runs out of 5. The cost is that the whole body sits in context on every run. Do not check any of this at home. Your own usage history ranks your skill first and it keeps its description. Ours passed on my machine and failed on a clean one. It worked on my machine. That sentence has been wrong for my entire career. #AI #ClaudeCode #AIAgents by @jackrudenko
0
15 days ago
Download
Our workflow already runs at Fable 5.5 level. Everyone else gets there in six months, when the model ships. We got there without a new model. What we changed was which model runs which agent, and how hard each one is allowed to think. Our agent workloads came out about 40% better. That is two model generations of improvement, from settings. Here is where it came from. When Fable 5.1 landed we ran our whole development workflow through the eval set. Every model against every effort level, agent by agent. The reviewer, the architect and the debugger needed Fable with room to think. The docs writer and the stack detector did the same job on Sonnet at low effort. The main session sits on Opus at medium and holds the rest together. We run seventeen agents, and almost every one wanted a different setup. Claude Code has no shape for this. A subagent model is static, written into the file once and frozen there. Effort is not a per agent setting at all, just one level per model for the whole session. So we rebuilt that layer in Magus. An agent points at a tier now, not a model. A preset maps four tiers onto a model and an effort level. Ours is called Opus with Fable help. Nine agents on smart, two on default, five on cheap, and one keeps its own setup because there is always one. One switch moves all seventeen at once, and the workflow, the plugins and the subagents stay exactly the same. Six months is a long time to wait for a number you can configure today. Claudish is next, so external models join the same tiers with their own effort rules. #AI #claude #DeveloperProductivity #AgenticAI by @jackrudenko
5
16 days ago
Download
The secret to working with Kafka is to think less. Anthropic tested it on their own models last week. It works. They published how each model does at every effort level, one task at a time. On a task about a slow Kafka service, both models got worse the more they thought. I haven't tried this approach with our Kafka engineers yet. It matches what we see in our evals at MadAppGang. Speed and architecture tasks suffer the most from extra thinking. My guess is simple. To make something faster, the model first has to decide what "faster" means. More effort gives it more room to decide wrong, with great confidence and a lot of tokens. Legacy code is its own story. A VBA port scored zero at every setting. I find that comforting, because some things in our industry never change. And when the model already knows the answer, max effort is a taxi that drives around the block before it stops at your house. "So should I just run everything on low, Jack?" you'll ask me. No. Security work and bugs you can reproduce gain a lot from effort. The point is to test effort on your own tasks instead of trusting the average. Swipe for every number. Data from Thariq Shihipar at Anthropic. Full article on the MadAppGang blog. #AI #ClaudeCode #AIAgents #LLM #DeveloperProductivity #BuildInPublic The secret to working with Kafka is to think less. Anthropic's own effort data, task by task. by @jackrudenko
1
17 days ago
Download
Twenty-five years ago, I started telling people that meetings hurt engineering performance. Now I'm the one adding them back, and I can't argue with my own reasoning. Three ways to work with an agent, and they produce three different engineers. On the left, there’s spoon-feeding: a small task, read every line, then the next small task. These are the engineers who tell me the cheap Chinese models match Fable, and for tasks that small, they're right. It says more about how they work than about the models. On the right, there’s letting the agent read and decide on its own. Reading a diff properly is hard, the agent gives the impression that it can judge its own output, and going along with that is the easy path. We call it meat proxy mode. OpenClaw and Hermes are built on this idea, and they're strong products. Software development isn't ready for it yet, except in a few narrow cases. In the middle is real delegation. The agent writes the code, so the human owns the acceptance criteria and the solution. That was always the hardest part of the job. Now it's the whole job. Two years ago, that work took two hours a week. Today it takes four to eight hours a day. That's the position that burns people out, and it's also the correct one. FOMAT doesn't help. Nobody wants to be the engineer who missed the window, so nobody stops on their own. So we ration it: one or two hours of hard thinking first thing in the morning. Then swimming, a walk, a tutorial, and meetings. One more hour in the afternoon, then stop. Two hours a week became three hours a day, and three is already too much. Most engineers burn out two or three times before they take a limit seriously. #AI #claude #FOMAT by @jackrudenko
0
22 days ago
Download
Amazon put the same coding agent on 50 teams and watched them for a year. Half came in under 3x, half hit a median of 4.5x, and the agent was not the difference. Amazon Stores watched 50 teams for most of last year. Normal mix of juniors and seniors, all of them working in old code, 90% on the same agent. They measured deployment velocity to production, so actual shipping. Half the teams came in under 3x. The other half reached a median of 4.5x, and a few passed 10x. Then Amazon went and interviewed them. The fast teams had changed how they work. The slow teams kept their old process and added an agent to it. We're about a year in with six teams across three companies, and the split looks the same to me. Here's the part I had backwards. I thought the fast teams started fast. Almost every fast team at Amazon reported the opposite, and their productivity dropped first. Two months of work inside the codebase came first. Better error messages, so the agent could see why it failed. A structure an agent could actually move through. Some teams moved Python and JavaScript over to TypeScript and Rust. A compiler talks to an agent better than a stack trace does. Our fast teams paid the same two months. Try explaining that invoice to a client who can't see the velocity yet. A team that skips those two months is still in the slow half a year later. I haven't seen an exception yet. What is hapening in your team? #AI #aiengineering #aitransformation by @jackrudenko
0
23 days ago
Download
Anthropic's own course said my system could not work the way I built it. � Their Claude Code Academy material on subagents says a subagent only gets the skills you explicitly write into its markdown file. � Magus does the opposite. A discovery agent picks the right skills for the task and hands them to every other agent in the flow. Almost nothing is hardcoded. That discovery is the whole reason it is useful. � So either the course is wrong, or my architecture is. � "Why not just believe the docs and rebuild it, Jack?" you'll ask me. And it's a great question. � Because rebuilding that flow with hardcoded skill lists is weeks of work, and it removes the part that makes it work. Before I pay that price, I measure. � So I wrote scenarios. A subagent with no skills in its file. A task that only finishes if the agent finds a skill first. Three passes each, because one run tells you nothing. � The subagents found the skills and used them, exactly like the main thread. The information in the course is not correct. � The real limit is somewhere else. In "claude -p", the non interactive mode, subagents are not available at all. That part was already known, and it is why magmux exists. It runs and controls any CLI tool in interactive mode, so the flow keeps its subagents. � This is what I spend money on. � Two to three thousand dollars a week, running experiments in madbanch. One run is a real Claude Code session with a controlled setup, real tool calls, and checks at the end. Every scenario runs three times. The score is a number between 0.00 and 1.00, because most agent work is partly right and a red light throws that away. A full suite is 360 runs, 24 minutes, and about $125. � Most of those runs tell me something did not work. The tool is slower. The setup is worse. The clever idea does nothing. � That is the useful part. Negative results are the cheapest knowledge I buy. � The same harness is now running the ten most popular agentic memory layers for Claude Code against each other. Same repo, same questions, same checks. The only variable is the memory layer. Results in a separate post. �#AI #evals #claude by @jackrudenko
1
24 days ago
Download

Download All

Save the photos and videos currently loaded on this page.

0Photos
0Videos
Ready to download 0 / 0

Keep this tab open until your download is ready.
back to up