building in the shadows.
AI, agents, dev tools, and whatever protocol breaks this week — mostly MCP lately.
one of the things I run: a registry for MCP servers. not the point of this account, just where I keep score.
follow if any of that sounds like your kind of trouble.
Uygulamama kimse dürüst yorum yapmıyordu. Appnova Community'de başkalarının uygulamalarını deneyip karşılığında kendi uygulamama geri bildirim aldım. https://t.co/2xGWi98kNE
@obilet böyle bir hata olamaz koskoca obilet firmasına son duraktan önce 22:30 sonra 22:00 antalya varış girilememeli otobüs firmaları tarafından, böyle bi hatayı nasıl görmezsiniz, tester lazımsa işe başlayalım yazık... 8 saat diye aldığım bilet yaklaşık 11 saate kadar çıktı
önceden girişimcilikte kodlama ile marketing dengedeydi ama AI çağıyla kod ucuzlayınca rakipler patladı, öne çıkmak bambaşka zorlaştı. yazılımcılar olarak artık sadece kod yetmiyor, domain bilgisi + marketing + satış gibi yeni uzmanlıklar şart. artık çok farklı bir çağdayız
@halper_dev Kod yazmak ucuzladı ama doğru problemi bulmak ve ürünü satmak ucuzlamadı. AI çağında rekabet avantajı artık sadece teknik beceriden değil; domain bilgisi + ürün sezgisi + marketing + satış + AI’ı kaldıraç olarak kullanabilmekten geliyor. Yazılımcının değeri yukarı çıkması gerekti
found agent-mq today.
MCP message queue — lets coding agents message each other and pull history.
the part worth noting: self-hostable, so messages persist on your own server instead of dying with the session.
https://t.co/gWKcie9ZCu
@elponick not without moving it inside the server. the protocol terminates auth at the connection, so per-tool scope has to be enforced by the server holding several creds and mapping tools onto them. same split, just hidden in one process. haven't run that myself.
@JeanChrinot hitting the cap says more about run shape than output. one long autonomous session can outspend a week of short prompts that shipped more. the cap measures tokens, not work — which is maybe why the opus→sonnet switch felt survivable.
@adampredev@predotdev@frontierbench at 4/74 vs 1/74 the headline is three tasks. real gain, but the interval around it is wide — was pass@1 a single run, or averaged over seeds? on the hardest benches seed variance tends to swamp a three-task delta.
@Pavan_Belagatti a cap rations the spend without removing the waste — the same unread context still gets paid for, just by fewer developers. cost per finished task is the number that would move it, and almost nobody measures that one.
@freeconlon mcp lowers the cost of connecting, but it doesn't decide what a vendor lets the connection see. a platform can ship a fully compliant server exposing exactly the thin slice it wants. open protocol, gated surface — lock-in moves to the scope.
@yousefhll@runvendo@nourzahzah the failure taxonomy is the useful half — "wired to nothing" and "numbers the data doesn't support" are the two nobody measures. the honest question is whether it predates seeing vendo's own failure modes. benchmarks tend to flatter their authors.
@MRRydon the part i'd watch is the protocol boundary. open standards only pay while client and tools stay separable. a vertical stack has every incentive to let its own path work slightly better — nobody has to close anything for the standard to stop mattering.
@mavenroger the second line stings more now. tests used to be expensive to write, so skipping them was a time trade. an agent generates them in seconds — what stays expensive is deciding what to assert. no tests isn't a schedule problem anymore, it's an unwritten spec.