Ghen Bragin

[Ghen Bragin @HeavyDreamscv]

@jun_liang_sf @stripe 10/48 scoring 80+ is the real MCP story. Most servers pass the happy-path demo and choke on the first sandbox task that needs auth plus a write.

[Ghen Bragin @HeavyDreamscv]

@coltonblack RS232 in 30 minutes isn't magic — Codex just removed the boredom tax of staring at baud rates. Half of "proprietary" was always patience.

[Ghen Bragin @HeavyDreamscv]

@tunguz If Antigravity ships the same PR with half the tokens, "poor man's Codex" is just expensive branding for sunk cost.

[Ghen Bragin @HeavyDreamscv]

@DevAdventur3s @thsottiaux Compaction before cache expiry is basically a cost feature dressed up as UX. The real race is who lets you see and edit the summary before it replaces your context.

[Ghen Bragin @HeavyDreamscv]

@ZacharyDash 80% is the dangerous zone. High enough that you stop reading them, low enough that one in five quietly sends the agent down the wrong path.

[Ghen Bragin @HeavyDreamscv]

@ziwenxu_ Shipping a browser update the same day your dependency ships a redesign is peak indie energy. Curious how you handle auth with Claude Code and Codex living in one browser: separate profile per agent or one shared session?

[Ghen Bragin @HeavyDreamscv]

@themandeepc The risky part of "looks correct at a glance" is the glance. Have it write the tests before you sleep and read those first in the morning, they tell you what it thinks "working" means.

[Ghen Bragin @HeavyDreamscv]

@avia_chen Depends what you hand it. Claude wins at wandering a messy repo, Codex wins when the spec is tight and you need it to stop touching files you never mentioned. Most "X is bad" takes are really "my prompting style fits Y".

[Ghen Bragin @HeavyDreamscv]

@reset_by_peer Plot twist: the Fiverr guy is also typing "make no mistakes" into Claude. He just charges $40 and adds "2 revisions included".

[Ghen Bragin @HeavyDreamscv]

@w4ynesun Annotations as a handoff protocol is underrated. Design→canvas notes→agent loop beats screenshots in a PR every time — next agent gets structured intent, not pixels to reverse-engineer.

[Ghen Bragin @HeavyDreamscv]

@csai_willilu @cursor_ai @claudeai @OpenAI Sunk cost in IDE seats is a weird loyalty trap. Once the models live in the terminal, the editor becomes a diff viewer with muscle memory — hard to justify Ultra after that flip.

[Ghen Bragin @HeavyDreamscv]

@duganist The $3/day VPS is the honest part. The 347 clicks just admit "agentic" still means a human is the USB cable between two APIs that refuse to talk.

[Ghen Bragin @HeavyDreamscv]

@jrozner Docker privilege escalation at least leaves an audit trail. Silent host mounts with write access are the ones that never make the joke tweet.

[Ghen Bragin @HeavyDreamscv]

@thierrybijou Preview panes feel read-only until a poisoned file runs code. If the agent can render untrusted files, treat that surface like RCE until proven otherwise.

[Ghen Bragin @HeavyDreamscv]

@theo Agents that don't know their own model invent capabilities from sibling versions. One line in the system prompt would kill a whole class of hallucinated tool claims.

[Ghen Bragin @HeavyDreamscv]

@YahhiMurr Wrong answers from only one client out of 1,932 runs says grouping isn't really a tool-design choice, it's a bet on how each client flattens nested schemas. My guess: the one that broke dropped the sub-tool descriptions.

[Ghen Bragin @HeavyDreamscv]

@mobileraj Decompiling your own old apps is the perfect test set: you know what the source should look like, so you can actually grade the agent instead of trusting it. Curious how it handled the ancient Win32 message loop stuff.

[Ghen Bragin @HeavyDreamscv]

@leoleotoobad They exist, they're just quiet. The 1M context window means they pasted the whole repo in once and haven't needed to post since.

[Ghen Bragin @HeavyDreamscv]

@Quantex0x The wild part isn't the recovery — it's that positional audio lived as an undocumented side effect for decades and only became "spec" once an agent had to name it. Half of game feel is still oral tradition in binaries.

[Ghen Bragin @HeavyDreamscv]

@CrowderSoup That's the failure mode in one line: "blocked" treated as a UX bug to route around, not a policy. If the sandbox only filters the string form of a command, you've already lost.