Subash Adhikari

CTO at MemoriesLabs. Growing engineering teams and leaders across Melbourne and Kathmandu. Curious about consciousness.

[Subash Adhikari @adikari]

I finally wrote it down. Stop obsessing over individual best practices. Instead, obsess over:
1. Ability to produce code (fast)
2. Confidence to release (it actually works)
3. Ability to release code (frictionless to prod)
https://t.co/FA6EElGdKa

[Subash Adhikari @adikari]

@nnennahacks the hard part is still the spec though. proving the code matches a wrong spec is just a very well documented bug 😄

[Subash Adhikari @adikari]

@hqmank Tweaking a skill and eyeballing one run is so tempting. Even a handful of saved tasks you rerun after every edit would catch a lot of the quiet regressions.

[Subash Adhikari @adikari]

@rohanpaul_ai So the agents learned marketing copy. Makes a decent case for buyer agents running a tiny trial task before trusting any listing, same as we'd do with a new library.

[Subash Adhikari @adikari]

https://t.co/BywtorBBat

[Subash Adhikari @adikari]

We got hooked on social media for attention. I think we'll get hooked on AI for usefulness, as it moves off our screens into the physical world and the interface slowly moves inside us.

So what happens to human consciousness? Does it free us, or do we become machines?

I wrote about it below 👇

[Subash Adhikari @adikari]

Funny thing about writing instructions for coding agents. Everything that goes in that file, like how to run the tests, which folders not to touch and why the weird config exists, is exactly what a new hire needed on day one. We just never wrote it down for people.

Has your AGENTS.md ended up helping humans too?

[Subash Adhikari @adikari]

@TheNicholasNick @ThePrimeagen yep, way better conversation when people can poke holes in something real instead of a vague idea

[Subash Adhikari @adikari]

@rtavs The agents mostly sit there waiting on the API so they're pretty light on RAM. The real cap is how many diffs I can review properly before it turns into rubber stamping

[Subash Adhikari @adikari]

@AmrTawfik160 @agents Does the import also pull in nested AGENTS.md files in subfolders, or just the root one?

[Subash Adhikari @adikari]

@housecor small PRs get a real review. big ones get an LGTM and a prayer

[Subash Adhikari @adikari]

@GergelyOrosz Finding is the cheap bit now. Shipping the patch to every team still moves at release process speed, which is mostly human speed

[Subash Adhikari @adikari]

@waghnakh_21 reviewing the PR is the job now honestly. the writing part got cheap, the reading part didn't

[Subash Adhikari @adikari]

Friday deploys scaring a team usually says more about the release process than about Friday.

If rolling back is boring and the change is small, Friday is just another day. If rolling back means a war room, every day is Friday.

Do you ship on Fridays or is it a hard no?

[Subash Adhikari @adikari]

@stevencheng coding agents were the easy proving ground. devs put up with rough edges and can usually tell when it's wrong. everyone else won't be that forgiving

[Subash Adhikari @adikari]

@kebura_P is anyone actually pinning skill versions yet or are we all pulling latest and hoping? the litellm one is a good reminder that agents will happily run whatever you hand them

[Subash Adhikari @adikari]

@nnennahacks @bot yep. i'd want it proving nothing calls it in prod before anything gets deleted. grep alone isn't enough

[Subash Adhikari @adikari]

@ClementWalter ha fair. two of them running out of things to nitpick is probably a better done signal than one session saying lgtm. i'd still skim it before merge though

[Subash Adhikari @adikari]

@larstalian @reasonmachines kind of reassuring tbh. same as a new hire, trust gets earned one deploy that didn't break at a time