Schwentker
@schwentker
Builder & AI education 🏗️ @at_proto #mpp 6 series #BlueskyCommunityVoices podcast w/ @bluesky pre-launch. @BlockchainU founder. xPayPal 📍 Oakland
A yr ago "agent" meant chatbot calling an API. 65 teams in one day moved the definition: a loop w/ a conscience. It plans, acts, checks work against the world, accepts being told no. Not louder models. Better loops. @awscloud.bsky.social AWS Builder Loft, San Francisco
5/5 StageLoop: @AshokSravanam gives an agent a listing URL + $5. It stages furniture in 3D, critiques itself & loops 3/10 → 8/10. “Loop until score ≥ 7 with zero geometric faults.” @akashnetwork.bsky.social zero.xyz devpost.com/software/sta...
4/5 AEGIS: Wally Ahmed built an on-call SRE that accepts “no.” Restart denied by a real @pomerium.io HTTP 403 → rollback → recovery in 24 sec w/ 0 human actions. “Every action must pass through a zero-trust policy gate it cannot bypass.” nexla.com zero.xyz devpost.com/software/aeg...
@manassharmahere 3/5 FleetYield: @manassharmahere & Sanskriti Uma made reliability a scheduling primitive. On 24 GPUs: +45.8% net goodput failures 3→0, premium-GPU misuse 32→0. “Treating failure risk as a first-class price reclaims the middle of the fleet.” devpost.com/software/fle...
2/5 melkat: Katerina Tchilinguirov & Melany Macías handed Claude $5 of real USDC & said: !!buy the Salesforce Tower!! 🏦 : )) It shopped a machine-to-machine market, built product, opened shop, cashed out real $2 Amazon gift card. @anthropic.com zero.xyz devpost.com/software/mel...
1/5 Enrichment Diet by Aristarkh Manchuliantsev Agent learned to stop paying. $0.43 → $0.14/candidate, −67%, w/ quality at 93/100. “Real-dollar feedback changes how you design agents.” Built w/ @akashnetwork.bsky.social @pomerium.io zero.xyz devpost.com/software/enr...
"You get a bad actor, good agent. You could have a good actor, bad agent." Mastercard's Craig Gilbert mapping the fraud matrix agentic commerce just added a dimension to.
"If I'm an average consumer, who's gonna make that right?" @Experian's Kathleen Peters on what happens when an agent transaction goes wrong & nobody's sure who's liable. Still no clean answer.
"We need one standard that everyone can collaborate around." Blackhawk Network's Nik Sathe on the AP2/UCP/ACP standards pile, and why merchants can't afford to guess wrong about which one wins.
"Capability doesn't translate to trust. Trust is earned, not scaled." @auth0byokta.bsky.social Abhishek Hingnikar watched an AI agent die in a password recovery flow live on stage, then said the line that summed up the whole night.
"A bigger change in commerce than when smartphones put the internet in everybody's pocket." @trySkyfire co-founder Craig DeWitt @CryptoCwby what's at stake w/ agentic commerce, from a room full of ppl trying to solve trust before the volume hits
Identity isn't authority. Neither solved yet for agents. Srini Vadhri @trulioo building agent certification w/ card networks. Question every network asks: "How do we know the agent?" Status: very early. Credential proves who sent action. Not who's responsible.
Where's an LLM even allowed to stand in a purchase? Badi Azad @PayPal - assistant completing buy inside itself "disassociates buyer from seller," strands merchant machinery built over a decade. Tax, fraud, callbacks, upsell. On agent identity & authority: "we're not there yet."
"When model becomes commodity, context is comparative advantage." Adeel Ali founder/CEO APIMatic.io Agent in versioned payments API w/o grounding invents namespaces, ships checkout marking orders paid w/o capturing money. Fix isn't a smarter model. It's context as contract.
& the @thesysdev team for showing up deep: @rabi_guha @ParikshitGD @zahlekhan @ritvikbudhiwhat w/ @ummunaz generative UI in builders' hands all weekend
thank you to the judges who gave builders real signal 🙏 not just scores: @DhruvAtreja1 @bigmarvin @mnvsk97 @killianm @ojusave @quarktosh @karsinkk @lotte_verheyden sharp questions, generous time
Sentinel: audits vendor marketing claims vs public evidence, writes machine-readable verdicts to cited.md & charges agents to read them. "marketing inflated for humans is invisible to agents." so build the layer agents can buy.
MCP Server Auditor: 6 prober agents sweep an MCP server for prompt-injection & scope-violation classes, publish cited verdicts & sell them over x402.org security for the tool surface other agents depend on, before someone gets badly burned.
Weather-Alpha: runs a physics weather model vs live prediction markets, trades wherever they disagree >8%. Kelly-sized, every 5 min, no human. forecasting agents don't ask to be believed. they post a position & let the outcome settle it.
Theta Desk: an options desk that wakes itself, screens 4,000 contracts in SQL & sells a $0.01 risk brief to another bot over x402 one agent makes a paid product. another agent buys it. the agent economy in miniature, running.
Carter Busse CIO @workato.bsky.social "We make more mistakes than the agents do. That's the real truth." -- on why Claude became first-line-of-defense. It holds institutional memory - 50 conversations back on Slack, fields filled in a year ago. Humans are worse at remembering than we think.
Barr Moses Founder & CEO Montecarlo Data "Data has been lying to us for years. It's just been doing it silently." -- on why agents lying w/ confidence & citations triggers more scrutiny than the quiet corruption we tolerated for decades. Psychology, not tech, is the constraint for autonomy.
@workato.bsky.social Rahul Dureja's healthcare POC: two versions of scheduling agent. 1rst exposed raw tools, hallucinated specialties, booked appointments w/ no dates. Same model. 2nd version: intent-based orchestration. Identical behavior every run. Architecture, not intel, is the variable
Okhtay Azarmanesh, @Snowflake "The single-agent ceiling. We already hit it." --on why multi-agent frameworks became urgent. Then the best insight (imho): "The collaboration, the team playing, is the gap." Models are brilliant individually. Clumsy together.
@adamse.bsky.social @workato.bsky.social CTO "They can draft an amazing email. But they can't deliver it. They can't get feedback. They can't coordinate & can't take action in systems on your behalf" -- on why smarter models alone don't cross reasoning frontier. Trust is hard problem underneath
"I'm unlearning everything I did over 12 years as a software engineer." @openclaw maintainer, GitHub HQ. Craft didn't improve. It got reassigned.
"I think you found Linux of AI." — @davemorin.com to @steipete.me Agents as websites. We're in 1997. Era of agents = as big as era of apps.
"It's not pull requests anymore. It's prompt requests." -- @steipete.me Every issue = signal someone's unhappy. Route it. Don't fight the slop.