Ilya Zel'

[Ilya Zel' @x777genius]

@nvandewetering Or route by role: Fable only as lead and reviewer, Opus or cheaper models doing the implementation. That's how I stretch limits in Agent Teams: https://t.co/nQQnyIJ2IU

[Ilya Zel' @x777genius]

@dianne_alter @conductor_build Same reason I wanted a workspace over raw CLIs. With Conductor winding down, Agent Teams might be worth a look: free, local, works with the coding agents you already use, plus a kanban the agents move themselves: https://t.co/nQQnyIJ2IU

[Ilya Zel' @x777genius]

@abhijitwt Half of babysitting is just checking whether it's done or stuck on a permission prompt. A desktop or Telegram/Slack ping fixes that part: https://t.co/sWxvTeXPG2

[Ilya Zel' @x777genius]

@adxtyahq Running 10 sessions overnight is fine, merging their output without a review gate isn't. Even an automated line-level pass on every PR in CI would catch a lot of that: https://t.co/n6Vyp4ivS6

[Ilya Zel' @x777genius]

@eng_khairallah1 Nice guide. If people want the same pattern without wiring it by hand: a strong model as lead, cheaper models as builders, a shared kanban and peer review, in a free desktop app: https://t.co/nQQnyIJ2IU

[Ilya Zel' @x777genius]

@WhaleInsider Rationing is why I stopped running everything on one frontier plan: a strong model leads and reviews, cheaper or free models do the grunt work. Agent Teams even ships a free no-signup model to start: https://t.co/nQQnyIJAys

[Ilya Zel' @x777genius]

@marfinxx The harness matters more than the model here: separate roles, another agent reviewing each task, and a board you can actually see. That's the setup in Agent Teams: https://t.co/nQQnyIJAys

[Ilya Zel' @x777genius]

@neil_xbt That 12.9% tail is usually stuff a line-level pass could flag before round one. I run AI review inside GitHub Actions on the subscription I already pay for, so every PR gets it: https://t.co/n6Vyp4hY2y

[Ilya Zel' @x777genius]

@sermakarevich The setup that worked for me: one stronger model as lead and reviewer, cheaper ones as builders, all on one kanban they move themselves. Open source if useful: https://t.co/nQQnyIJAys

[Ilya Zel' @x777genius]

@Kanojiyaaakash1 The setup that worked for me: one stronger model as lead and reviewer, cheaper ones as builders, all on one kanban they move themselves. Open source if useful: https://t.co/nQQnyIJ2IU

[Ilya Zel' @x777genius]

@DailyDoseOfDS_ Discovery is half of it. The other half is keeping the same skills installed, updated and cleanly removable across every agent you use, and that's the part I automated: https://t.co/k5JhIy0q7Q

[Ilya Zel' @x777genius]

@Corporate_CJP Not dumb at all: writing got cheap, verifying didn't. What helps me is an AI pass with line-level, severity-rated comments on every PR in CI first, so the human review goes to design and intent: https://t.co/n6Vyp4ivS6

[Ilya Zel' @x777genius]

@AdarshWritess For me the biggest jump wasn't picking one, it was running several together: one plans and reviews, the others implement, all off a shared kanban. Built an open-source desktop app for exactly that: https://t.co/nQQnyIJAys

[Ilya Zel' @x777genius]

@Layton_Gott Cheap models like this shine as builders: let a stronger model plan and review while the cheaper ones implement in parallel. I run that split as one agent team on a kanban: https://t.co/nQQnyIJAys

[Ilya Zel' @x777genius]

@DN_degen The annoying part with all these setups is keeping them in sync across every agent you use. I add the skills once and let a CLI write each client's native format, with a lockfile for the team: https://t.co/k5JhIy0q7Q

[Ilya Zel' @x777genius]

@navi_Ai2 @Teamily_AI Taking turns is the bottleneck. Once agents claim tasks off a shared board, message each other and review each other's work, you only step in when something's blocked. That's what I built, free and local: https://t.co/nQQnyIJ2IU

[Ilya Zel' @x777genius]

@RoundtableSpace Workflow skills like /spec and /plan are exactly what you want identical in every agent you use, not copied by hand per tool. One command installs the same skill across Claude Code, Codex, Cursor, Copilot CLI and others: https://t.co/k5JhIy0q7Q

[Ilya Zel' @x777genius]

@haider1 Honestly the limit resets matter more to me than the launches now. I stopped leaning on one frontier model for everything: a strong one leads and reviews, cheaper or free ones do the grunt work in the same team: https://t.co/nQQnyIJAys

[Binance @binance]

A vulnerability (worse_query) affecting iOS 26.7.1, 27.0.1, and 27.2 allows malicious apps to read local files from other apps, putting seed phrases, private keys, and login credentials at risk across the entire ecosystem. This is an iOS-level issue, not unique to Binance.

Immediate Action Required:
- Delete any apps installed from untrusted sources (such as sideloaded or manually signed IPAs).
- Update iOS and all apps to the latest official versions.
- If you suspect a breach, transfer funds to a new wallet created on a clean device and reset all active login sessions.

Never install unauthorized packages, and keep your software up to date.

[Ilya Zel' @x777genius]

@jussisaur That's the case for putting a reviewer on even the 1-line bumps: it can catch a latent bug on main that a human would never bother to exercise. I run that in CI with my own models: https://t.co/n6Vyp4hY2y