Vicky Prasad

Agents and the evals that keep them honest. 6 yrs shipping payments, wallets & mobile apps in production. TypeScript · Go · React Native.

[Vicky Prasad @prasadvicky231]

hanvil won @hedera's Improve the Hedera Harness prize at @ETHGlobal ETHOnline 2026 🏆

A full Hedera network in one 10 MB binary, 5.8 MB RAM. The official local node: 18 containers, 3.8 GB.

Coding agents run the harness offline now. Zero testnet HBAR.

cc @hedera_devs

[Vicky Prasad @prasadvicky231]

@thdxr a subagent can call it done once its tests pass, even when the test is the thing it changed. who reruns main's checks before something lands in ready for review, the gooey session or you?

[Vicky Prasad @prasadvicky231]

@simonw the older issues come from substack's undocumented archive api. if that one keeps running next to the rss import, how do you find out when they change it

[Vicky Prasad @prasadvicky231]

@sydneyrunkle @link @stripe if the user approves but the response gets lost and the agent retries, is that a second charge or the same order

[Vicky Prasad @prasadvicky231]

@thymikee what if it gets stuck at 0.03 and never reaches 0.02. does it keep trying forever or is there a limit

[Vicky Prasad @prasadvicky231]

@jarredsumner the new bun check already runs into it. in a repo with a "check" script it ran mine instead, and -p . didn't change that. only reached the type checker from another folder. intended, or should the builtin win?

[Vicky Prasad @prasadvicky231]

@sh_reya do all 1,772 tags come out the same with sharing on and off? a 4b model tagging long traces seems like it could flip a few just from different batching

[Vicky Prasad @prasadvicky231]

@omarsar0 consistent can also mean wrong the same way every run. a judge that misreads one tool call misreads it every time and the scores look stable. what's it checked against, human labels on a slice?

[Vicky Prasad @prasadvicky231]

@kacperkapusciak @TesterArmy what about the collisions that aren't on screen. two agents on one test account, one deletes something the other's flow needed. does that show up in the report as a bug, or do you catch it?

[Vicky Prasad @prasadvicky231]

fixed a re-render in our RN app that fired every 8 sec. two things broke

the "4m ago" labels and a useMemo missing a dep. both were running on the bug

[Vicky Prasad @prasadvicky231]

@mattpocockuk How does it decide a line is a no-op? Some rules only matter on the one task a month that hits them. Does it look at past sessions or just read the file?

[Vicky Prasad @prasadvicky231]

@Hiteshdotcom What took the sting out for us was shipping JS fixes over EAS Update. Review only gates native changes after that. Is this one stuck on a native change or a first submission?

[Vicky Prasad @prasadvicky231]

@mrousavy Read the perf doc. Thread CPU and Android jank are in, JS profiling is deferred. When the JS thread is the hot one, what does the agent go on? Pull a Hermes trace with release-profiler, or just read the code?

[Vicky Prasad @prasadvicky231]

@rauchg Peeked at the bundle. The walkthrough code is strings the model wrote, not read from the files. When it edits models.remote.ts later, what stops the two drifting apart?

[Vicky Prasad @prasadvicky231]

@chrispader That was quick

[Vicky Prasad @prasadvicky231]

@jpudysz Yeah the Android numbers surprised me, closer to StyleSheet than iOS. Thought it'd be the other way round.

[Vicky Prasad @prasadvicky231]

@ismine91 @reactnative @wcandillon @chrfalch @expo Is the Picture built in a worklet on the UI thread, or on JS each frame? Wondering what a GC pause does in the middle of a fight.

[Vicky Prasad @prasadvicky231]

@mattpocockuk crap-score works with Vitest, just tried it. It reads istanbul's coverage-final.json, so add json to Vitest's coverage reporters and point npx crap at the file. Docs only say Jest. Gating CI on it or just a report?

[Vicky Prasad @prasadvicky231]

A one-line change broke every build of a React Native app.

await import(spec), with the path built at runtime.

The comment said Metro would skip a computed import. Metro fails the whole bundle instead.

Typecheck green. Tests green. CI never built the bundle.

Does yours?