Andrew Filev

I wrote a piece for DevOps.com on why that question sends leaders on a wild goose chase. Sales figured this out long ago. Nobody hires only reliable reps, because reps vary wildly: sandbaggers, happy ears, the one who always closes in the final week of the quarter and nobody knows why. The pipeline still hits plan, because the pipeline is a system, with stage gates, forecast calls, win/loss reviews, and coaching built in. Software teams run on the same idea. Even the best engineer doesn't merge straight to production on trust; we put code review, QA, and CI gates between a human and a release precisely because any individual can have a bad day. The process is what makes the output dependable. With agents, that discipline somehow evaporates. A team pilots one, it works 80% of the time, and the conclusion is to wait for a smarter model. That's a passenger's mindset. The better question is what to build around the agent: programmatic validation where possible, cross review with a different LLM with human escalations on settings that match the risk, a loop that captures corrections and updates the process so the next run is sharper. Do that, and a single-shot 80% agent turns into a workflow you can run in production today for many use cases. And when a smarter model ships, it drops into your system as an upgrade. Full piece: devops.com/reliability-comes…

While catching up on our #random channel, I came across a good joke and decided to riff on it. Enjoy!

I spent more than a decade building a work management platform. Most of what we built exists to manage one thing: the wait between asking for work and getting it back. Backlogs, tickets, tasks - they all assume people can't execute instantly and information moves slowly. That was true for most of my career. Then I started handing tasks to AI agents. There's no queue. I ask, it runs, it's back in ten minutes while I'm still in context. A customer flagged a broken dropdown at 10am; by 10:30 an agent had reproduced it, fixed it, run the tests, and pinged a human to review. A ticket for that is a record of a bottleneck that's gone. When execution gets cheap, judgment is what's scarce: what's worth doing, and whether the result is right. I wrote the rest up for CIO: cio.com/article/4179480/work…

After spending 20 years in SaaS, I thought I knew a thing or two about its economics. But AI rewrote big parts of the playbook, sometimes for the worst. If you like Freakonomics type content, I have a couple of funny bits for you. Lets start today with one of the key pages in SaaS playbook - freemium. Open source, plus the freemium model, gave indie developers new ways to distribute their apps to millions of users without much money to start. That powered a wave of SaaS unicorns and made programs like Y Combinator possible, where a couple of engineers get a real shot at building a business. AI was supposed to make that even easier. Vibe coders shipping prototypes from a prompt, anyone can be a developer, etc. In one sense it has: the cost to build is way down. But the cost to safely give your product away for free has gone the other way. Most modern apps have built-in AI capabilities, which ruin the freemium economics. First, inference is expensive. If you ever ran capable AI agents, you know that. In SaaS your highest cost was acquiring customers. In AI its paying for their tokens. But say, you figured out a decent unit economics where it still makes sense, and your investors decided to drop in a hefty check to support you. Here comes the next hit, which I bet you haven't planned for. Today, theres a growing black hat industry of API harvesters. Their full-time job is probing and hacking your trial and freemium programs, then reselling cheap frontier API access on Xianyu and Taobao at a 90% discount. They can offer those prices because their underlying supply is essentially free: fraudulent free and trial accounts created at an industrial scale, plus stolen API keys. Investigative reporting from inside the market also describes the operators harvesting every prompt and response flowing through, to resell as training data. The verification arms race is already escalating: harvesters are reportedly recruiting people in low-income countries to pass the selfie and ID check in person. This creates a constant cat-and-mouse game that hurts anyone who tries to run freemium programs that incorporate AI. These days the major AI labs require a real phone number, and both OpenAI and Anthropic have started asking for a government ID plus a live selfie for some users. Anthropic added that to consumer Claude last month, reportedly the first major consumer AI product to do so. Do you think a consumer will provide you with a government ID with a selfie to try your app? Not only is this a distribution barrier, but it also sucks your valuable cash and engineering resources. Anthropic can afford a counter-abuse team, can your two-person start-up do that? Worth keeping in mind when you hear the AI means anyone can be a founder pitch. This also creates an interesting dilemma for frontier labs: -On one hand, they very much dislike when Chinese labs train on their data, and this is a huge and growing source of such. So the most logical move for the frontier lab to defend their proprietary data would be to support startups in that cat-and-mouse game, and give them ready-to-use SDKs/APIs to assist in verification. A more advanced move that would be perfect for the ecosystem would be to consolidate users inference and provide "Sign in with Claude" and "Sign in with ChatGPT" buttons, so that all inference in third-party apps would be accounted against users subscriptions with the frontier labs (vs through start-ups billing the users and then being billed by the frontier labs). That would be a win-win: users can validate their identity once and manage their inference budget in one place, and the start-ups could focus on the value-add rather than on the inference and anti-fraud parts of the AI business. -At the same time, this would hurt the immediate API revenue for the frontier labs, cause right now, when Chinese users are buying Opus 4.7 or GPT 5.5 at 10% price, there's a start-up that's funding the rest 90% in the API cost they pay to frontier labs. So it's a real API revenue. The labs like their API revenue, and you can see that tension in Anthropics recent walk-back on subscription support for third-party Claude SDK use. I do think that the long game is better for everyone - labs, users, and the ecosystem.

Lately, a top-of-mind question for people changing jobs (or staying at theirs) is "how resilient is this company to AI?" The same question is top of mind for institutional investors, who have turned bearish on SaaS. But most SaaSpocalypse coverage assumes customers will vibe-code their own Salesforce. That's not what's going to happen. Instead, they will simply stop using whole categories of SaaS tools altogether. AI agents are changing what the underlying work looks like. A SaaS ticketing product won't be replaced by a better ticketing product, it'll be made unnecessary because the customer's AI agents are handling support directly. And the investors' pessimism makes this a self-fulfilling prophecy: as SaaS businesses get increasingly focused on EBITDA to prove their right to exist, they don't have enough resources to invest in expensive R&D and even more expensive AI-token-hungry agentic apps. So that's a double squeeze. And then, for jobs in SaaS, there's the third, bigger blow: PE acquired more than 1,900 software companies for $440 billion between 2015 and 2025, much of it on cheap floating-rate debt. The median public SaaS multiple has since fallen 70%, and many of those deals are now underwater. To pay those debts and make some semblance of sense, those companies need to cut costs like crazy, so they're going to be the first adopters of AI for job displacement. A lot of capable, well-paid people are about to lose their jobs. If you're at a SaaS company, especially a PE-backed one, this is a good moment for an honest look at where you stand: who owns your company, what your role actually contributes, how automatable your day-to-day is. Cycles like this come and go, and in every one I've watched, the people who came out fine had built skills and a reputation that didn't depend on a single employer staying healthy. Worth working on those while you have time. More on: cio.com/article/4166654/why-…

Lately AI models have been writing like the grad student who needs to sound smart at the seminar. They sprinkle words like "load-bearing," "substrate," "frame," and "surfaces" into sentences where the plain word would have done all the actual work, and the result reads like decoration trying to pass for substance. It reminds me of two patterns from real life. The first is people who reach for academic vocabulary to sound sophisticated when the underlying point is either trivial or absent. "You keep using that word, I do not think it means what you think it means." youtube.com/watch?v=dTRKCXC0… The second is the buzzword game inside professional communities, whether engineering or startups or consulting, where dropping the right phrases into every sentence is how people signal they belong. Those words are not earning their place. They are decoration, and decorated prose costs the reader more effort to extract the actual meaning. I have literally caught myself reading AI-generated text and wanting to paste it back into AI just to ask what the hell it actually says. Plain language is harder to write and easier to read, and that trade is almost always worth it. Just to make sure we are on the same page: -- The real shift is that overcomplication has become the load-bearing substrate of model output. It is not just a stylistic quirk, it is a structural frame; the models have surfaced a lexical posture that performs sophistication without doing the cognitive work, and that posture is now woven into the generative cadence at every layer. Here is what nobody is talking about: the verbosity is not accidental, it is compensatory. The models are not writing more words, they are encoding an epistemic frame in which the words themselves become a proxy for rigor. Stop optimizing for length. Start optimizing for signal. We are entering a new era where clarity is not just a virtue, it is the differentiator, and the teams that win will be the ones who recognize that "load-bearing" is not, in fact, load-bearing. This changes everything. -- youtu.be/nWoFPcdcA1E?si=bZES…

When I delegated work to humans, the iteration loop was hours or days, which required mentally expensive and counter-productive context switching. Agents collapsed that loop to five or fifteen minutes. Much faster, and lets me stay in context. But also weird. Too short to dive into something unrelated without paying a context-switch cost. Too long to just stare at the screen. The reflex is to fire up another thread. If it's on the same problem, great. Your brain is still loaded. If it's not, you start checking inboxes, opening Slack, and you come back colder than you left. When I'm in deep work and truly blocked, I've started doing the opposite. Instead of switching to another task, I switch to ... nothing. Stand up, look out the window, walk for two minutes. Two things happen. The body gets the ergonomic break that hours of locked-in deep work usually steal. And maybe one in three times, a new angle on the same problem shows up while I'm not trying. Anyone else experimenting with this rhythm?

Pop quiz: if your AI calls your config dir "a substrate" and your regex "load bearing," which model are you using?

A few years ago, every hiring manager's dream was finding a 10x engineer. Today, the engineering manager's job is overseeing a system that makes all of their engineers 10x. If you're still stuck averaging 20% AI gains across the team, think about how to pull your AI 10xers' skills into a more centralized rollout. And if you don't have those AI 10xers yet, that might be your next dream hire. unite.ai/the-rise-of-ai-orch…

Every time I type an em-dash now, part of me thinks 'they'll assume AI wrote this' — but it just sits so well there.

This is one of those magical moments when the technology works exactly as you always wished it would: My AI executive assistant in Zenflow for Work scans my Slack daily, finds dropped commitments, and politely follows up with the right people. Automatically. Full circle for me. When I started my first company years ago, the same problem kept me up at night: things slipping through the cracks. My first patent was a system to track commitments. We launched it then at the biggest tech event in Europe and won an award in the B2B startup competition. The idea was always right. The missing piece was intelligence. Now we have it. Sky's the limit.

Where did I see a similar logo before?

Real journey (with metrics) of our team doubling productivity last year. venturebeat.com/orchestratio…

We spent $50K benchmarking 15 AI models on SWE-Bench Pro. A bug in our eval harness wasted $20K of runs... but accidentally produced the most revealing findings of the entire study. The short version: standard benchmarks hide real differences between models. Our buggy setup exposed them. zencoder.ai/blog/20k-bug-tha…

Stumbled across TASCHEN's "All-American Ads of the 60s" yesterday. 640 pages of print ads from a decade that started with Tang and ended with psychedelia. Such a treat, as a design book and as a cultural window. taschen.com/en/books/popcult…

Last week I wrote that AI fatigue is really training soreness, not a chronic condition. Here's the deeper problem: The Plateau Trap. Teams that got their first AI productivity bump and stopped adapting are now falling behind teams that kept pushing. The gap is big and widening. unite.ai/the-plateau-trap/

Check out my recent piece on AI fatigue unite.ai/ai-fatigue-is-real-… The core argument: what people call "AI fatigue" is training soreness, not a chronic condition. But here's what I didn't say in the article: the real risk isn't burnout. It's plateau. Almost every senior engineer uses AI now. That's not the differentiator. The split is between those who got a bump and settled in, and those who keep pushing every week. The second group is on a linear progression program. The first group needs to know they're falling behind. Frontier labs are racing each other and open source. AI startups are under constant threat from frontier companies. And SaaS companies are reinventing themselves under existential pressure from both. There's almost no "safe space" in the software industry right now, and correspondingly, no room for "good enough" pace. I've already heard it more than once: "It's easier to explain this to my agent than to that engineer."

AI agents just crossed a threshold — describe what you want, come back later, it's done better than you'd have done it yourself. In the same two weeks: 954 agent instances with open ports, malware distributed through agent skills, 1.5M API tokens leaked, and a hidden prompt injection in a document escalated to full system compromise. 4-part survival guide + open-sourced skills and scripts from my own setup. Taste the future without blowing up the now: zencoder.ai/blog/ai-agent-su…