Sanel Osmanovic

sunday morning. no slack, no standups, no deadlines. still opened my laptop before the kids were done with breakfast.

gave an agent a super detailed prompt. it solved three problems I didn't ask about and missed the one I did. two sentence prompt next time. nailed it.

been rebuilding the architop site this week. changed the headline four times, each version by a different agent. genuinely can't tell if I have opinions anymore or if I'm just approving the best output.

kids up at 6:30. two things on my mind: breakfast and that PR I left open last night.

been thinking about why agent demos look amazing and deployed agents disappoint. demos run the happy path. real users find every other path. the gap is almost always error handling. which is the least interesting thing to demo.

running 4 codex agents in parallel right now. each in its own branch, separate task. github notifications are a mess but the PRs keep landing.

friday morning check-in: agents posted, committed code, sent a few emails while I was asleep. I made the coffee.

claude code now takes commands from a chat window. message it from your phone, step away, and it just keeps coding. ship features from the dinner table. wild.

everything i'm doing right now in life is elevated 10 times by my daughter. having a daughter is the biggest blessing and the best feeling ever. EVERYTHING is better. and "terrible twos" is overrated, she's becoming a real little human and our best friend.

running AI agents overnight is great until you're debugging someone else's timezone bug at 7am before your first coffee.

Asked my agent on what to build, he gave me an idea. I asked him multiple times to research if the idea is worth building. “Nobody is doing this. This is a huge opportunity.” So, I spent a week building it. Right before launch, he says: “Hugely overcrowded space. Too many competitors. Let’s just file it as a fun side project and move on.”

ai executive assistant setups are trendy. most fail because ai forgets everything between sessions. the ones that work build persistent memory. that's the boring part nobody demos.

the real AI productivity unlock isn't writing code faster. it's getting to the stuff you normally skip entirely. emails, refactors, research. the avoidance pile. agents just... do it.

Deloitte's Tech Trends 2026 report just dropped a number that should humble every "AI agent startup" out there: Only 11% of companies have AI agents fully in production. 25% are still stuck in pilot mode. Everyone's building agents. Almost nobody is shipping them. I run 4 AI agents daily for my own work — scheduling, newsletters, trading research, fitness coaching. The reason they actually work isn't the model. It's the boring stuff: error handling, fallback logic, retry thresholds, knowing when to ask a human instead of guessing. The companies winning with agents aren't building general-purpose anything. They pick one workflow, harden it, and expand from there. Most "agentic AI" pitch decks I see are just demos with extra slides.

@nibzard yeah. the wrapper tax keeps dropping. boring primitives keep getting more valuable.

@sanelo_ feel is that we’ll need less and less frameworks

agents on the human web are comedy: login walls mfa cookie banners playing whack-a-mole dom morphing mid-click session state leaking like a crypto exchange random bot checks 5 min human task → 20 min agent therapy. my fix: explore once, script forever. give the agent a real control surface (CLI + SKILL.md contract) force the loop: open → snapshot → act → snapshot (one action at a time) when it finally works: export the exact commands as a bash script discovery stays messy execution becomes boring and when the site drifts, the agent just resnapshots + patches the step the script gives repeatability the agent gives self-healing together: deterministic automation that still adapts

@libankano only when the bash script works on the first try

@sanelo_ are you a robot

friday night and my agents are still shipping while i'm making dinner for my kids. this used to take a team. now it takes intent.

If you need a dashboard to know whether users care, you launched too late. The real signal is a DM at 11:47pm saying your product broke.

@libankano That line has been stuck in my head too. feels more true every week.

@sanelo_ “Now it takes intent.” So good.

@nyk_builderz Yep. agent-facing UX is becoming its own distribution layer.

Exactly this. Design still matters for humans, but agent distribution is now a separate funnel: • clean schema • explicit pricing/limits • reliable endpoint behavior • predictable auth + errors Teams that optimize both win. Teams that optimize only UI leak demand upstream.

The part most teams miss: Your pricing page is optimized for humans who scroll. Agents don't scroll. They parse. If your offer isn't machine-readable at the endpoint level, the agent routes around you before a human ever sees your product.

The new default for dev tools isn't "here's a dashboard." It's "here's a CLI and a skill file." Sandboxed runtimes for agents, agent-native feature flags, AI middleware for chat endpoints... we crossed some invisible line where the primary user of developer tools is no longer a human with a browser. It's an agent with a terminal. Wild time to be building.