iiinova — where your digital self keeps evolving

Schwentker @schwentker.sandboxlabs.ai · 20h

9/9 Arty Vohmincevs & Aaron Brown @mercor on what makes agents different: "They don't tire, they don't sleep, they don't ask, they just go." If agents are the third customer, who is a platform really designing for?

000 0

Schwentker @schwentker.sandboxlabs.ai · 20h

8/9 Tobi Lütke @tobi, Shopify: "We are now all verification engineers. We are building harnesses, we are building acceptance tests."

100 1

Schwentker @schwentker.sandboxlabs.ai · 20h

7/9 Sarah Sachs @sarahmsachs, Notion on an agent that opened a browser to type into Notion like a human after a failed edit: "Models don't give up. And that's not always good."

100 1

Schwentker @schwentker.sandboxlabs.ai · 20h

6/9 Katelyn Lesse, @anthropic.com on agent memory: "Dreaming is when you look back on all the past sessions and aggregate new memories that you add into memory..."

200 2

Schwentker @schwentker.sandboxlabs.ai · 20h

5/9 Diogo Almeida @CompleteSkeptic @typesafeai.bsky.social on a voice demo built w/ Jev: "I was like holy crap, that's the missing thing to get rid of keyboards entirely."

100 1

Schwentker @schwentker.sandboxlabs.ai · 20h

4/9 Assaf Elovic @assaf_elovic, Ora, on agents & the open web: "We spend 20 years keeping bots out and now we have to let them in." Btw agents reach llms.txt in only 32% of cases & 86% of those b/c the site linked it.

100 1

Schwentker @schwentker.sandboxlabs.ai · 20h

3/9 Olawale Oladehin, AWS, on the speed of the shift: "Just six months ago, 3% of Vercel deployments were triggered by coding agents. Now that number is 60% in six months."

120 2

Schwentker @schwentker.sandboxlabs.ai · 20h

2/9 Guillermo Rauch @rauchg, Vercel CEO, on the numbers: "The number of deployments they trigger with coding agents has grown 28 times. The haters said AI would take your job. But instead, you gave AI a bigger backlog."

320 3

Schwentker @schwentker.sandboxlabs.ai · 15/10/2026

4/4 Wallets, mandates & card agent programs exist. What I tried to combine: drift detection, versioned rules, human approval & rollback. All at The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

000 0

Schwentker @schwentker.sandboxlabs.ai · 15/10/2026

3/4 Self-hosted LiveKit on Fly, real-time voice agents on a tiny CPU, & why the agent proposes but the money stays in code.

100 1

Schwentker @schwentker.sandboxlabs.ai · 15/10/2026

2/4 Hard ceilings it can't raise, versioned rules in @mongodb.bsky.social Atlas, voice approval via LiveKit & ElevenLabs, auto rollback if the new deal costs more. Testnet, made-up articles.

100 1

Schwentker @schwentker.sandboxlabs.ai · 09/10/2026

5/5 The sound stack: phone line (@twilio), ears (streaming STT, e.g. @DeepgramAI ), brain ( @cresta Conductor), voice ( @elevenlabs.io ). WebRTC via @LiveKit can save ~300ms. Past ~300ms, silence feels unnatural. So: when you pause mid-sentence, how does a machine know you're thinking, not done?

000 0

Schwentker @schwentker.sandboxlabs.ai · 09/10/2026

4/5 Next for Veronica: plan your team’s night out. 12 SF startup events in. Plain code applies the hard rules ($40, Thu, fits 8, near SoMa) & 3 survive. Natali’s Jev classifier ranks them, picking only from that list, so it can’t invent an event. Code filters. Jev picks.

100 1

Schwentker @schwentker.sandboxlabs.ai · 09/10/2026

3/5 The payments plan: an AI agent w/o a spend policy books a kayak trip & gets charged 10x in pathUSD on Tempo.xyz testnet. "Veronica's" refund tool is a verb, not a pen: refund_overcharge(charge_id), not send(to, amount). The model can't invent a payee.

100 1

Schwentker @schwentker.sandboxlabs.ai · 09/10/2026

2/5 "Veronica" answers the unBoring Club line. She checks who's calling, finds the charge, files a dispute w/ a ref #, & hands off to a human if needed. Built on Cresta Conductor: 1 umbrella agent that spins up sub-agents (Verify, Charges, Refund, Close). Agents building agents. ☯️

100 1

Schwentker @schwentker.sandboxlabs.ai · 01/10/2026

7/7 Jevathon afterparty on the @coderabbitai roof deck: DJ Lidia, the Bay Bridge & @HKrackDev on sax, solo, w/ an "Ode to Jev." Thx to @typesafeai & @AICollectiveCo & everyone who built w/ Jev.

000 0

Schwentker @schwentker.sandboxlabs.ai · 01/10/2026

5/7 "Optimization is already good with the numbers, but it misses the meaning embedded in how people describe their requests," wrote @tanishkgovil. His Jev-fed hospital elevators cut emergency waits from 46s to 13s across 16 held-out scenarios (self-reported).

100 1

Schwentker @schwentker.sandboxlabs.ai · 01/10/2026

3/7 "Optimization is already good with the numbers, but it misses the meaning embedded in how people describe their requests," wrote @tanishkgovil. His Jev-fed hospital elevators cut emergency waits from 46s to 13s across 16 held-out scenarios (self-reported).

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/09/2026

6/6 @typesafeai.bsky.social grew Jev on synthetic data into a System 1 model: fast, typed judgment, no prose. Image input is next on the public roadmap. After that, maybe a System 2? Hackers are chomping at the bit. When judgment costs fractions of a cent, what's left to decide slowly?

000 0

Schwentker @schwentker.sandboxlabs.ai · 27/09/2026

5/6 Thx to organizers @AICollectiveCo, @AJs_AI , @roangws, @chappyasel, My Luu, plus host @coderabbitai.bsky.social Shoutout to its intern trio behind Revolution, 1st place: a wildlife ecosystem sim running ~4,000 Jev decisions for ~$1.30.

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/09/2026

4/6 "We woke up a night later with 1,500 pull requests open, and not enough humans in our company to press the button." @HKrackDev of @coderabbitai.bsky.social on unleashing background agents on a monorepo. Answer: Triage ranks agent PRs by risk & review-readiness.

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/09/2026

3/6 Agents tend to think it's more expensive than it is." Alex Warren ( @exrhizo.bsky.social ) of @typesafeai.bsky.social re: Jev pricing. 720 states x 30 Qs? Likely <10¢. Tip: Jev always answers, so expecting a non-answer = type error. Add a "not applicable" choice instead.

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/09/2026

2/6 "If it's vibes, now we're getting somewhere." @allietheicon.com of @typesafeai.bsky.social on Jev's sweet spot: pure numbers belong on a CPU; semantic judgment belongs in Jev. Btw, Jev isn't agentic. Does nothing until code puts it in the loop. ( w/ @stephencouncil )

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/09/2026

5/5 My bet for a @StripeDev meetup a year from now: shopping agents, research agents & seller agents buying & selling over MPP w/ stablecoins. We'll be asking what they bought, what they paid & who gave them permission.

000 0

Schwentker @schwentker.sandboxlabs.ai · 24/09/2026

4/5 “It's enforced at the infra layer.” AWS's Chethan Shriyan on agent spending limits. Hasan Tariq's demo wallet held nearly $1M in testnet funds. The agent's session budget? 15 cents. Wallet balance isn't spending permission.

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/09/2026

3/5 “It's like you do your homework & then you rank yourself & grade yourself.” @DeividasMat on one AI doing all the shopping jobs. @m11labsai's OpenMarket splits them: 5 sellers, 1 buyer & a referee checking the claims.

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/09/2026

2/5 Lunch w/ @collision. A meme-filled one-pager for Patrick. Weeks later: Stripe.directory. @MbyM on making businesses searchable for people & agents, down to the CLI. “I'll do it today.” No Q3 2027 backlog.

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/09/2026

2/5 Lunch w/ John Collison. A meme-filled one-pager for Patrick. Weeks later: Stripe.directory. Stripe's Mark Moriarty on making businesses searchable for people & agents, down to the CLI. “I'll do it today.” No Q3 2027 backlog.

100 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

7/7 @sofiiiiiasz emcee: On how AG-UI first got shown at a WorkOS meetup w/ a 5 min demo slot, & where it went from there: "You never know where it'll lead you." Good reminder that these places especially in SF 🌉 are where standards often start Thx hosts CopilotKit Manufact @workos.bsky.social

010 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

6/7 Andrew Khadder, Manufact Ran his whole event ops thru ChatGPT Voice. T-shirt orders, allergy lists, Slack, waitlist approvals, expense report. "I didn't need to build out a whole dashboard interface" The dashboard didn't vanish. It got decomposed.

100 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

5/7 Tyler Slaton, @CopilotKit Best framing of the alphabet soup all night: "AG-UI is basically like you give all of these tools as Lego blocks" & the agent decides which blocks, in what order. The frontend isn't a screen the agent talks about. It's part of the agent's operating environment.

100 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

4/7 @idosal1 co-creator of MCP Apps & AG-UI, live from Tel Aviv @ some ungodly hour Planning a trip = 10 tabs. But you don't need 10 whole apps, you need a room selector, a map, an approval card. "You break it down to the atoms." An app stops being a destination & becomes a set of capabilities.

101 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

3/7 @tobinsouth.bsky.social The supply chain problem MCP has that APIs don't: "you can just change the tool list at any time" @tyler_cpk's analogy harmonized: "imagine if your Node packages could change under you w/ no version pin. That's roughly what we've accepted."

110 1

Schwentker @schwentker.sandboxlabs.ai · 20/09/2026

2/7 @tobinsouth.bsky.social, @Anthropic Most uncomfortable idea of the night: (Claude Code) "auto mode reduces risks over humans who generally just click yes" Asking permission at every step looks safer. But people stop reading by prompt 20. User focus moves to budgets, scopes & thresholds.

110 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

11/12 Skeptics Jev from @typesafeai.bsky.social is a classifier. Zero-shot, many criteria per call, calibrated. Typed output, logprobs, temp scaling, thresholds in app code: none of that is new & objections may hold. What differs is where calibration comes from. Training objective, not later step

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

10/12 Attorneys Jev from @typesafeai.bsky.social isn't an LLM. It classifies against criteria set in advance & returns a value w/ each disposition. Prompts leave no record of a standard applied. A threshold w/ version & date does. Producing a standard is one posture. Reconstructing it's another.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

9/12 Researchers Jev by @typesafeai.bsky.social isn't an LLM. It classifies against set criteria & returns a full distribution, not just a top label. Same winner, two meanings eg: spiked @ 0.91, flat @ 0.19. Flat says the state lacks the answer. Calibration holds across groups, never one answer.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

8/12 Investors Jev from @typesafeai.bsky.social isn't an LLM. It's a judgment layer under one: fixed question, typed answer, confidence attached. Generation shipped first. That layer still gets hand-built at every company, out of parts built for other jobs. Rebuilt separately, where bugs live.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

7/12 Regulators & Evaluators Jev from @typesafeai.bsky.social isn't an LLM. It works alongside one at the decision points, classifying each case against criteria set in advance. Same test every case. Each result carries a confidence figure logged at decision time, not reconstructed later.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

6/12 Board & CXOs Jev from @typesafeai.bsky.social isn't an LLM. It answers questions & reports how sure it is, as a number. That number draws a line eg: approve @ 0.8 & up, lower to a person. Today that line may sit in a prompt. A threshold can be reviewed & defended. A paragraph cannot.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

5/12 Tech Leader Batching 13 checks into one prompt looks efficient until answer 3 starts conditioning answers 4 thru 13. Arithmetic, not a context limit. Hard to catch, since contaminated answers may look like clean ones. Jev evaluates each against same state in parallel. A 14th costs ~nothing.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

4/12 Non-tech Builder Prompt-heavy apps tend to put us in every loop. One reads the reply & decides. Jev works differently. A builder writes the question once. The answer comes back as a value plus a number, eg: yes at 0.53. A rule sits in software: below 0.7, hand to a person. Above, let it run.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

3/12 Engineer Jev returns typed answers plus calibrated confidence. No text generation. Many independent questions per call. Same label, two branches. Jev doesn't pick one. A threshold in app code would & it can differ eg: 0.6 for a read, 0.95 for a write, same call. Thresholds become reviewable.

100 1

Schwentker @schwentker.sandboxlabs.ai · 16/09/2026

2/12 Layperson Jev answers a question one sets in advance & tells how sure it is. Ask a language model whether a msg is a refund request & it generally writes a paragraph explaining itself. Ask Jev & it returns yes, plus a confidence score eg: 0.53. Both may say yes. Jev would say how close.

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/08/2026

Mind-blowing deets from @ rayyanzahidai says runs agent swarms across ~30 machines: • Token budget: ~50 BILLION tokens/month • Smallest agent: Runs lightweight at ~2,000 tokens • Largest agent: Heavy lifter running 30M–40M tokens in a single context window

000 0

Schwentker @schwentker.sandboxlabs.ai · 27/08/2026

Agent swarm scales w/o limit. Oversight doesn't. It collapses to one point: a master agent, one human glance standing between infinite parallel work & what actually shipped. Real limit isn't how many agents run. It's how much a person can still verify as scale grows.

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/08/2026

Inside Rayyan's agent prompt architecture: • Dynamic agent scaling tied directly to task dependencies • Strict Ralph loop verification & 0 stub data • Deep research 1rst via CDP & Browseruse • Smallest deployable units using git worktrees + @cloudflare.social CLI • Indy execution on unblocked tasks

100 1

Schwentker @schwentker.sandboxlabs.ai · 27/08/2026

Agent swarm masterclass @cloudflare.social Immersive Commons hackathon by @rayyanzahidai The vision: spawning multi-agent teams using @claudeai Code to handle complex tasks w/ max effort & min manual intervention. Only the master agent requires HITL oversight.

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/08/2026

8/ i ponder: as this rail gets built, what would it take to make sure it works as well for the smallest publisher as it does for the biggest platform

100 1

Schwentker @schwentker.sandboxlabs.ai · 24/08/2026

7/ Parag: "We aren't very opinionated about the (payment) rails." Today @p0 "Index" pays content partners "the boring way," invoicing. But he's clear where it breaks: "you need to pay the very long tail of all content owners in an efficient, simple, standard payment rail." #x402 #MPP

100 1