Yogesh

So Something

[Yogesh @yog_codes]

@NeverSinkDev mcp on an image editor is the useful bit. most agent image flows just fire edits and hope. if the tool hands the rendered file back into context so you can catch a wrong crop, thats when it stops being a gimmick

[Yogesh @yog_codes]

sophos + openai Daybreak for cyber MDR

agents cut avg investigation from ~38 min to ~89 sec (96%). ~52% of MDR cases resolved end to end. destructive actions still need a human

@OpenAI
https://t.co/EZeBMYL4Tc

[Yogesh @yog_codes]

@mr_kozh yeah this is underrated. every half baked shortcut left in the repo becomes the next agents example. cleaning that up before you let it loose is basically part of the harness now

[Yogesh @yog_codes]

@JoePro checking where a memory came from before it sticks is the right gate. once a page plants an instruction it outlives the tab, and thats a nasty failure mode for long running agents

[Yogesh @yog_codes]

@rohanpaul_ai held out logs as the check is the useful part. one catch from backend land, old logs from before an api or tool change will still score fine while the skill teaches a call that doesnt exist anymore. would version tag the logs and only score against the current tool schema

[Yogesh @yog_codes]

@arantespp agents forget, graphs dont is the right split. having the agent read whats ready before it starts beats hoping session memory survived the last crash

[Yogesh @yog_codes]

@doodlestein yeah the credits make the cloud path tempting. still feels like the second you need weird ports or a fat local checkout, people bounce back to their own harness

[Yogesh @yog_codes]

@simonbs remote mac + install link in chat is the missing piece. annoying that the embedded preview breaks on codex remote though

[Yogesh @yog_codes]

@tushaarmehtaa arcade as a benchmark is kinda clever. watching an agent lose at pacman tells you more than another leaderboard screenshot

[Yogesh @yog_codes]

claude haiku 5.5 just dropped

built for the cheap/fast lane: summaries, classification, coding subagents, browser use. ~75% cheaper to run than haiku 4.5 on avg, and its the first haiku with adjustable effort

api: claude-haiku-5-5 @AnthropicAI
https://t.co/ogtM5BYPU0

[Yogesh @yog_codes]

@Luxiontx 11k for a film studio skill is a lot. the hard part isnt the first render though, its whether the agent can redo one bad shot without restarting the whole timeline

[Yogesh @yog_codes]

@iamlukethedev talking to word/excel through an agent is usually where these demos get awkward. if hermes can rewrite a doc and fix formulas without bouncing apps thats the useful part

[Yogesh @yog_codes]

@athrix_codes @droid @FactoryAI oneshot is great for the hero section vibe. the second you need a real feedback loop though (tests failing, design tweaks) the iterative harness still wins

[Yogesh @yog_codes]

@fatih @cursor_ai this is underrated. treating agent workflows like something you actually revise each week instead of just stacking more prompts. the transcript dump for the retro is a nice trick

[Yogesh @yog_codes]

@mitalasam 42 diagram types is a lot. half the value is catching when it picks the wrong one and quietly drops a branch from the flow

[Yogesh @yog_codes]

@rimtoln the lead + specialist split is usually where single chat setups fall apart. once you can see who owns which block, a bad plan is way easier to kill before it floods the whole run

[Yogesh @yog_codes]

@sirshibaninja @atomicagent_io local + free + cross platform is what people keep asking for. glad they split desktop from the cli so terminal folks dont get a surprise ui

[Yogesh @yog_codes]

@marouane53 this is the kind of skill that actually helps. reverse image alone lies a lot so stacking vision + signs + light is the right call. proof trail matters more than a confident guess

[Yogesh @yog_codes]

@rishit30g @claudeai @AnthropicAI useful. does the pace warning fire off rolling usage over the window, or only once you cross a hard % of the 5h limit?

[Yogesh @yog_codes]

@yetone nice. when it scores a session heavy on tool results, does magpie just show the chart or does it suggest truncating that slice next run?