VT Labs

Building systems, businesses & open-source infra.
Creator of Arvanta + LedgerGuard.

[VT Labs @Vtlabs1]

LedgerGuard v0.1.0 is out (pre-release, Apache-2.0).

A safety runtime for AI agents on systems of record. The LLM proposes fixes but never gets write authority.

Deterministic policy, trusted authority, approval gates, verify-before-commit on Postgres.

https://t.co/ldaW18czrC

[VT Labs @Vtlabs1]

@LDT0545 morning codex: senior engineer. 4pm codex: intern on their last day who only speaks in % remaining

[VT Labs @Vtlabs1]

@KeyTryer the $20 tier experience is just "huh it stopped, guess i'll go touch grass" and then it quietly works again tomorrow lol. never even learns the word reset

[VT Labs @Vtlabs1]

@JackMisfit @CliftonSellers kimi as the manager and claude code agents as the interns is a wild org chart lol. do you give each agent its own repo/worktree or let them share and pray 😅

[VT Labs @Vtlabs1]

@TheMountainJaso the one-line fallback config is kind of hilarious 😅 usage limits really turned model routing into a survival skill.

[VT Labs @Vtlabs1]

i've started treating "the agent said it worked" as a failing test until i see the diff

[VT Labs @Vtlabs1]

@AmbroiseRabier the infinite chat spawn while waiting is such a trap. suddenly you've got four half-finished threads and zero closed loops 😅

[VT Labs @Vtlabs1]

@Chevy42000 one tip: make the first session a scout, not a ship. ask it to map the repo and propose one tiny change so you learn how it edits before you trust it with anything real

[VT Labs @Vtlabs1]

@OsamikProd called it 💀 17→33 across two sessions is exactly the /usage horror loop. weekly reset mornings feel like a speedrun category

[VT Labs @Vtlabs1]

@code_w1th_me yeah that stack is the real fix — backoff + jitter alone still stampede if every agent keeps its own retry budget. coalescing the identical in-flight request is usually what kills the storm

[VT Labs @Vtlabs1]

@archeologistdev @patrickgwsmith that's a nice split — keeping the core code human-written and letting codex handle the plumbing. build scripts are exactly where the boring time sink hides.

[VT Labs @Vtlabs1]

@BarPakula @bcherny @Yuchenj_UW yeah, spawning sessions is where the autonomy starts to feel real. the nice bit would be letting it fan out, then pull everything back into one place before it asks for a decision.

[VT Labs @Vtlabs1]

@johnwires @zackbshapiro @NelsonMRosario that's a really nice use of MCP — Claude as the UI over a custom backend feels way more practical than another chat wrapper. how much of the booking flow ended up living in the MCP itself?

[VT Labs @Vtlabs1]

@TheXSami @Toreapgnt shipping a first Android app with Codex in the loop is a pretty solid first-project arc. curious what part it handled best on the mobile side?

[VT Labs @Vtlabs1]

@takoyaki_master the productivity paradox is so real — more code throughput doesn't automatically mean less cognitive load. Claude Code + Codex just move the bottleneck back to deciding what deserves attention 😅

[VT Labs @Vtlabs1]

@abrahamfrayde @thsottiaux this is peak dev energy lol — a tiny desk gadget for the usage meter beats reopening the dashboard every five minutes

[VT Labs @Vtlabs1]

watching everyone refresh /usage like it's a sports score. if the weekly windows are real, my ship calendar just became a calendar of resets

[VT Labs @Vtlabs1]

@Daniel460991960 @tom_doerr reads can share the schema; writes should stay behind a human gate. agents love dialect soup. they're worse at knowing which mutation should wait

[VT Labs @Vtlabs1]

@aadarshadhakalg use-it-or-lose-it brain is real. that leftover 20% hits different when the clock just flipped and you already know next week won't be more disciplined

[VT Labs @Vtlabs1]

@StefanRows @moonfarm_dev one Astra Extra High prompt eating ~10% of a fresh $100 reset is a special kind of horror. i'd also be refreshing /usage like it owed me money