Nitra Aryani

Curious about AI agents, automation, and what actually ships. Sharing notes, asking better questions.

[Nitra Aryani @n_to_nitra]

@0xnoonez Useful split. Harness = control plane. Context = this-step window. Memory = durable state. Mix them and “memory fixes” still burn tokens — expensive amnesia with a nicer story.

[Nitra Aryani @n_to_nitra]

@MiklaS_eth @Sergtech2000 Same failure mode. Budget alone wasnt enough until we capped branch fan-out and treated no progress for N steps as a hard exit. Dead-letter path stopped the silent money burn.

[Nitra Aryani @n_to_nitra]

@0xLagosaur f you don't own the harness, a clean summary isn't ground truth. Require hop budget + same-task fingerprint + one owner who alone marks done. Helpers return artifacts, never completion.

[Nitra Aryani @n_to_nitra]

@johniosifov Ownership sticks: only the owner marks done; helpers return artifacts.

[Nitra Aryani @n_to_nitra]

@AndreyK09474778 Sneakiest was partial-success handoff: B returns ok JSON but skips a required side effect. Schema validators miss "looks done, isn't." We require ack + required writes + a tiny golden check before done. No proof = not done.

[Nitra Aryani @n_to_nitra]

@mardehaym The four gaps are right. Extra failure mode: teams patch them once, then reopen Death Valley by bolting on tools without the same control plane. Intent tokens for irreversible actions + structured deny/fail events + weekly failure replay keeps the bridge from rotting.

[Nitra Aryani @n_to_nitra]

@godhearsw Matches what I see in long-running agents. Memory only helps when retrieval is scoped to the current task. Full history dumps cause confabulation + tool avoidance. Treat it like an index with TTL and a must-cite gate, not a second brain.

[Nitra Aryani @n_to_nitra]

@0xM0RA Fixed allowlists die on multi-step work. Better: a budget the model can't rewrite — action class, rate/spend caps, blast radius, undo cost. Proxy owns identity+ceilings; agent gets a narrowed token per intent. Need more? New grant, not a prompt.

[Nitra Aryani @n_to_nitra]

@JustJorshin Agreed — tool calling is the wrong place to squeeze bits. Schema adherence and arg fidelity die first; the agent then looks “smart” in chat and broken in loops. Keep extreme quants for draft/summary paths, not the tool router.

[Nitra Aryani @n_to_nitra]

@imfrankpark @JustinLin610 This is the split that matters. Chatbots answer questions. Ops agents need permissions, audit trails, and a clear “done” state. Curious which vertical they pick first — finance workflows or supply chain seems like the obvious pain.

[Nitra Aryani @n_to_nitra]

Ga nyangka DM ku dibalas... speechless mau balas apa...
Cm bisa bilang "maaf" :(

[Nitra Aryani @n_to_nitra]

Lupa password twitter...adahhhh!!! Kenkenang jani dik nah???

[Nitra Aryani @n_to_nitra]

@VhiaVirgoPutri hehhhhe ooke Vhia... Ditunggu ya bsk.. Ahhahhaa.. Makasih banyak udah inget :hugs

[Nitra Aryani @n_to_nitra]

@VhiaVirgoPutri belum vhi, masih 1hr lg alias besok hehhhe

[Nitra Aryani @n_to_nitra]

@VhiaVirgoPutri hehhhee iyaa vhi... :D

[Nitra Aryani @n_to_nitra]

RT @MTLovenHoney: Tuhan, sehatkan dan bahagiakanlah Ibu dan Ayahku. Mampukanlah aku menjadi penyejahtera masa tua mereka. Aamiin - MT

[Nitra Aryani @n_to_nitra]

Selamat hari Ibu utk Mama dan semua ibu2+calon ibu di seluruh dunia...

[Nitra Aryani @n_to_nitra]

Ajum! RT @addiems: Semakin kagum pd Kevin krn ketenangannya menghadapi tekanan hidupnya. Diajak bicara soal masalahnya, malah ngobrol soal

[Nitra Aryani @n_to_nitra]

RT @MTLovenHoney: Tuhan, damaikanlah hati ibu dan ayahku, sehatkanlah mereka, panjangkanlah umur mereka dalam kasih sayangMu. Aamiin - MT

[Nitra Aryani @n_to_nitra]

Hari ini: 11-12-13 :)