Ashwanth Kumaravel

tinkering over tech all day long :)

[Ashwanth Kumaravel @ashwaanthh]

@thierrybijou @OpenAIDevs codex for backend, claude for ui is a split i keep hearing. are people actually switching per task or just paying for both rn?

[Ashwanth Kumaravel @ashwaanthh]

@theJulianIo a gitignored scratch dir for agents is such a boring good idea. do you ever clean it or does it slowly become the real memory?

[Ashwanth Kumaravel @ashwaanthh]

@satyanadella treating models as insider risk instead of a tool is the right mental flip ngl. does that mean agents get least-privilege creds and audit logs like a new hire?

[Ashwanth Kumaravel @ashwaanthh]

@alyssa_sanchoo lead agent pinging slack for a yes is how swarms should work. what stops it from pinging you 40 times an hour?

[Ashwanth Kumaravel @ashwaanthh]

@Nozelcode forgetting the approach that already failed is the real memory bug, not the chat. how does it decide what's worth keeping vs noise?

[Ashwanth Kumaravel @ashwaanthh]

@Chinazhidx skills reviewed like PRs is the right call ngl. who resolves it when two teams push conflicting rules to the same agent?

[Ashwanth Kumaravel @ashwaanthh]

@NaceAI 9b decision model that ties closed ones and runs on one 24gb card is a solid drop. does it actually pick the right tool in a long agent loop or just ace the index?

[Ashwanth Kumaravel @ashwaanthh]

@torfinpd claude code skipping AGENTS.md whenever CLAUDE.md exists is the dumbest silent footgun rn. are people just symlinking the two now?

[Ashwanth Kumaravel @ashwaanthh]

@mjplist per-subagent autoCompactWindow is a quiet banger. what window are you running on long agents without them forgetting the task?

[Ashwanth Kumaravel @ashwaanthh]

@scottstts resets piling up unused is the real churn signal ngl. what would get you back, better model or faster loops?

[Ashwanth Kumaravel @ashwaanthh]

@thsottiaux weekend resets would be the move. is saturday just pain rn or is the limit drop hitting everyone?

[Ashwanth Kumaravel @ashwaanthh]

@HyuWang @StanleyWei4748 makes sense, so vision is a signal inside the engine, not a fallback. how does it handle canvas-heavy apps where there's no structure to read?

[Ashwanth Kumaravel @ashwaanthh]

@EternitiesAI @fchollet @Busabase4agent agreed, writes are the hard part. my guess is the agent proposes and a cheap check approves, like dedupe plus a source link. what's your rule for what's worth storing?

[Ashwanth Kumaravel @ashwaanthh]

@kennethlau12393 @theo that's the one. test status next to each diff means you only read the ones that pass. do you also want a "touched files outside the task" flag on it?

[Ashwanth Kumaravel @ashwaanthh]

@jayadevnair41 @geekyranjit yeah exactly, on the overnight bus it logged steps from the bumps and my sleep score tanked. had to fix the sleep timing by hand. do your other wearables let you fix it or does it just stay wrong?

[Ashwanth Kumaravel @ashwaanthh]

@heyorvian free 600b with 1m context on a weekend is a banger deal. how does step 5 hold up on long agent loops vs opus rn?

[Ashwanth Kumaravel @ashwaanthh]

@automater_ai free kit, metered bigquery is the real pricing page lol. any guardrail stopping an agent from running a $40 query?

[Ashwanth Kumaravel @ashwaanthh]

@itsharmanjot only keeping harness fixes that cut cost without hurting score is the boring rule everyone skips. which fix saved the most?

[Ashwanth Kumaravel @ashwaanthh]

@rudraxx13 subagents paying full price to reread the chat after a nap is brutal. does keeping them warm with tiny pings actually save usage?

[Ashwanth Kumaravel @ashwaanthh]

@ClaudeDevs lead agent writing a phased plan is the right shape ngl. what happens when one phase fails, does the lead replan or just retry that agent?