Arcos

ur fav researcher ai and creator

dms open

[Arcos @0xarcos]

Companies pay $300K+ a year for engineers who can run AI agents without babysitting them

Last week 20 Claude Code agents shipped 143 changes while I slept. I approved 6 merges

Here is graph engineering in Claude Code, file by file. Free ↓ https://t.co/RF9Pw3N4Yc

[Arcos @0xarcos]

@Vlixsi exactly false negatives are the metric that matters

[Arcos @0xarcos]

Companies pay $300K+ a year for engineers who can run AI agents without babysitting them

Last week 20 Claude Code agents shipped 143 changes while I slept. I approved 6 merges

Here is graph engineering in Claude Code, file by file. Free ↓ https://t.co/RF9Pw3N4Yc

[Arcos @0xarcos]

@Vlixsi by what slips through. catch rate means nothing if bad changes still reach production

[Arcos @0xarcos]

@harleyfoote_ not all were wrong. most simply failed to earn enough evidence to merge

[Arcos @0xarcos]

@jacewbuilds changed interfaces reopen all dependent gates

[Arcos @0xarcos]

@zkjays yep, writing moved to agents. accountability didn’t

[Arcos @0xarcos]

@shashwatj_ failed tests, contract violations, conflicting diffs or weak evidence
the graph treats uncertain as no instead of pushing it downstream

[Arcos @0xarcos]

@vtc volume is only useful when rejection is cheap and explicit

[Arcos @0xarcos]

@Vlixsi good graphs optimize for rejection, not output volume

[Arcos @0xarcos]

@kylekane that’s the right gate for external actions. the graph can do the work, but publishing stays human

[Arcos @0xarcos]

@_alphashark_ the graph does
tests reject most, verifier agents review the rest, and humans only see the six that survive

[Arcos @0xarcos]

@TheEaglegon known behavior goes into deterministic tests, while a separate verifier attacks the diff with edge cases and requirement checks. if something still escapes, it becomes a new regression test before the next run

[Arcos @0xarcos]

@kylekane you don’t
that’s why trust comes from gates, not from assuming the unseen changes are fine

[Arcos @0xarcos]

@Zhen4u5q typed specs before execution, atomic diffs at every handoff, then an independent reviewer before merge. waiting until the final review is already too late

[Arcos @0xarcos]

@gaurav0bhowmick exactly. generation is cheap, trusted integration is the product

[Arcos @0xarcos]

@dong9xbg each agent gets its own branch and file ownership. if two tasks need the same core file, the graph serializes them or sends the conflict to an integrator agent

[Arcos @0xarcos]

@mayanksaxenaai exactly. judgment helps, but the final gate should be deterministic

[Arcos @0xarcos]

@fuelmeupcc that rejection rate is the point. the graph filters volume into signal

[Arcos @0xarcos]

@sabrinodirenzo the review graph is what makes that volume actually usable

[Arcos @0xarcos]

@crlazeeee the real shift is making the human the final gate, not the entire loop

[Arcos @0xarcos]

@crlazeeee thanks crlaze