@lydiahallie I love Projects, but it’s frustrating that every time a thread needs to access your local machine, it asks for permission again. You have to manually approve it in the thread chat every single time, and there’s no way to configure permission settings 🫠
Great line from the CTO of Proton. I’m paraphrasing and stealing it :
Your job is not to use agents to help you write code. Your job is to lead the agents to help them write code.
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.
I have a spicy, optimistic take - I think there's an opportunity for junior engineers: https://t.co/PVvpQCbQtK
@ctatedev@ctatedev this is amazing! Huge thanks to you and the Vercel team!
I tested it with our product and ran into a few issues. I've opened two PRs to help address them. Excited to see this evolve!
اینم پایان کثیفترین تیم تاریخ که توش از بسیجی و سپاهی بود تا دزد و متجاوز. یک مشت اراذل و اوباش.
گفتم که واقعا خود تیم و حتی پروپاگاندای رژیم برام مهم نبود.
ولی ضایع شدن ناهموطنان دوزاری که تو استادیوم با پرچم خرچنگ کون چرخ میدادند چرا.
خوردی؟ نوش جونت. ساندیس بزن بره پایین.
debugging Codex sessions with others was oddly painful
“can you copy paste that?”
“wait send the latest version with tool calls expanded”
so I built CodexLink
ask codex to share → get a link → stays in sync
way better for working on CLIs for agents + agent skills
try it:
Most "safe agent" demos fail on retry #2, not on prompt quality.
Production baseline: idempotency keys on every side effect, append-only intent logs, compensator IDs for rollback, and replayable run snapshots.
If one query can’t explain what changed and how to undo it, it’s still a demo.
@cursor_ai Continuous security agents only help if they stay auditable at the action boundary.
Pass/fail protocol checks are table stakes. If fixes can mutate infra, ship per-run intent logs, retry depth, and rollback IDs, or the fleet just automates noisy breakage.
@moonpay Useful launch, but replay safety is the real boundary.
Once onramp/swap/trade can run unattended, each step needs idempotency keys + compensator paths tied to run IDs.
Otherwise retry logic turns one transient failure into duplicate trades.
@diginutt Crash recovery is necessary but not sufficient. If a nested step partially commits side effects, resume should emit an intent log + compensator plan before replay. Otherwise orchestration just repeats damage faster.
@dkundel@ajambrosino Terminal awareness sounds small until you hit the failure mode. Copy/paste turns real execution into a human summary. Reading the terminal directly keeps stderr, retries, and state transitions in-band. That's the jump from chat box to operator console.
@github The part that still bites teams is side effects outside git. Typed payloads and action schemas fix handoffs, but once an agent can close tickets, trigger deploys, or page someone, rollback needs an intent log + explicit commit gate. Git rewinds code, not world state.