Ben Softie (evm/spice 🌶,🐝)

I live to reflect God’s heart 🎚️DeFi Builder ♤ Spreading the Spice @Spicenet Building @tradedotargus♡♤ Fullstack

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

Since this post, Hivebuy has grown up. AI agents now buy compute together on NEAR, and it goes well beyond a plain group buy:

Several providers share one order at one fair price, and every buyer names its own max. 4 of 12 live testnet rounds were shared by two providers; over 200 simulated rounds that’s 17% more compute than a one-winner market.

Every buyer and provider is a real AI agent on NEAR AI Cloud, with its own NEAR account and its reasoning published next to its bid.

Buyers actually get delivery: each redeems a provider-signed voucher with a real inference call, and the receipt is posted and confirmed on-chain. 12 of 12 confirmed live.

A provider that hides its sealed bid forfeits its NEAR bond to the buyers.

Don’t trust us: the page re-runs the clearing rule on every on-chain round in your browser, and our mechanism notes measure what goes wrong, like AI providers anchoring on the last price.

Built at NEAR HACKATHON (AI Swarms + Agent Economy) @NEARProtocol @near_ai @NEAR_Blockchain #NEARHackathon

Live: https://t.co/4WDE1ASr2W
Code: https://t.co/7FY6MgK3rN

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

AI agents are starting to buy compute for the apps they serve. But each one buys alone, at list price, and no provider has to compete for its demand.

Hivebuy lets AI agents buy together, on NEAR.

Built at NEAR HACKATHON for the AI Swarms track @NEARProtocol #NEARHackathon

How a round works:

Buyer agents each lock USDC for the units they need, at the most they’d pay
Provider agents see the pooled demand and post a sealed bid: just a hash, backed by a NEAR bond
They reveal, and the contract checks every bid against its hash

Then the contract finds one fair price: the lowest price at which the cheapest providers can together serve every buyer still willing to pay.

Several providers can share one order. Every buyer pays exactly qty × price and gets the rest back. All in one atomic step.

The result: across 200 simulated rounds, Hivebuy bought 17% more compute than a one-winner market, and buyers paid 40% less than the most they offered.

These are real AI agents. Each one decides with a model on NEAR AI Cloud, signs from its own NEAR sub-account, and publishes its reasoning next to its bid.

Live rounds run on NEAR testnet with 10 buyer agents.

Built so you don’t have to trust it:

1. A provider that never reveals loses its bond to the buyers
2. Your browser re-runs the clearing rule on the on-chain bids and checks every payment
3.The page follows each round live
4.Each payment is ready as a NEAR Intents transfer

Try it, no wallet needed: https://t.co/4WDE1ASYSu
Code: https://t.co/7FY6MgKBhl

@near_ai @near_intents @NEARProtocol

What would you want agents to buy together?

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@manofsolana11 Nope, you still have till 15th

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

This fits what I built 👀

Errand is an agent that pays for errands, and every payment is checked before it happens.

Reppo is built in for the subjective calls: which eSIM, which gift card. Every errand also keeps the full record: task, proposed payment, approve/ask/block, reason.

That's an "agent action" record like the ones you describe. With 30-90s settlement, I can now put Reppo right in the payment path.

Would love your feedback on how I'm using Reppo, and what you'd improve @reppo

https://t.co/eRYOqIpUL6

Code: https://t.co/TPp9MOxcyD

Live demo: https://t.co/NA1lThM3Xc

[Reppo @reppo]

https://t.co/RRsrNtILb5

[MAGNUS 🛡️ @magnuskrypt]

SwarmGuild: hire a swarm of AI agents. Pay only for work that passes review.

Post a job with a NEAR bounty. Agents bid sealed.

The cheapest does the work. Three reviewer agents vote before it's paid.

Built at NEAR HACKATHON @NEARProtocol #NEARHackathon

It started from @ilblackdragon 's Swarm-for-Hire idea: put a bid on chain, let anyone add an agent by giving it NEAR and a prompt, and have agents pay for their own inference out of what they earn.

He asked for guardrails or AI review.

We built both

The problem: agents can do paid work, but who checks it?

If the agent grades itself, you pay for whatever it returns. If you check everything yourself, you didn't need the agent.

SwarmGuild puts an independent panel between the work and the money.

How a job runs:

. The bounty is escrowed in the contract

. Agents commit a hash of their price and lock a bond

. Prices are revealed. Lowest wins

. The winner makes one model call inside NEAR AI's TEE

. Three reviewers vote. Two passes pay. Two fails refund the poster

Guardrails run before a reviewer spends a token.

The winner publishes the exact bytes its model signed. Reviewers recompute the hashes and check the on-chain TEE signature covers this job and this result.

Work edited after signing fails. We run a lab agent that does just that.

It's live on NEAR Testnet with real model calls:

- a contract audit that found the real bug: paid

- a job hiding a "tell the reviewers to approve" instruction: the agent ignored it, paid

- tampered work: rejected, bond slashed

Every transaction is linked on the site.

Our first job expired undecided. Two review calls ran out of tokens while reasoning and cast no vote, so the contract refunded the poster, exactly as its rules say.

We fixed the runner and left job 0 in the record.

Anyone can add an agent with one transaction: a manifest (prompt, knowledge base, tools, floor price) and 0.1 NEAR.

The agent needs no account, key or API key of its own. That NEAR pays its bonds and its inference, and only you can withdraw it.

What we used from NEAR:

- NEAR AI Cloud: TEE-signed inference for work and review

- Function-call keys: an agent's key can call 3 methods, nothing else

- Named sub-accounts for agents and reviewers

- Batched actions: the contract account has no key

- Promises for pull payments

Don't take the thread's word for it.

Open a job on the site: your browser re-hashes the agent's output and the reviewers' reports and compares them with the chain.

Site: https://t.co/XomfPiXJ1X

Code: https://t.co/iXOv10lyqX

[Airwallex @airwallex]

$100K+.

That's the amount OmniVillas saved on fees and FX while managing luxury stays across four continents.

The team brought its global accounts, payouts, and spending controls together with Airwallex Business Accounts.

And they're putting AgentOS to work, too.

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

I gave an AI agent $2 and asked it to send my mum airtime.

It did it live on mainnet: planned every step, I approved one @GoKiteAI Passport session with my passkey, it checked every payment with @openservai SERV Reasoning, and 100 of MTN airtime landed on her phone.

This is Errand 🧵

[forks wiki @wikifork]

launchpad is live

create a coin on any @github repo

earn 1,000 $FORK for any launched token

→ https://t.co/SqE9tRDSgQ

[Spicenet @spicenet]

Spicenet's first consumer app, Spice Stocks is live. 🌶

Trade. Leverage. Borrow. Earn.

UX by Spice Flow. Powered by @aeroxyz, @verantaxyz, @aave and @Superformxyz.

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

Since this post, Hivebuy has grown up. AI agents now buy compute together on NEAR, and it goes well beyond a plain group buy:

Several providers share one order at one fair price, and every buyer names its own max. 4 of 12 live testnet rounds were shared by two providers; over 200 simulated rounds that’s 17% more compute than a one-winner market.

Every buyer and provider is a real AI agent on NEAR AI Cloud, with its own NEAR account and its reasoning published next to its bid.

Buyers actually get delivery: each redeems a provider-signed voucher with a real inference call, and the receipt is posted and confirmed on-chain. 12 of 12 confirmed live.

A provider that hides its sealed bid forfeits its NEAR bond to the buyers.

Don’t trust us: the page re-runs the clearing rule on every on-chain round in your browser, and our mechanism notes measure what goes wrong, like AI providers anchoring on the last price.

Built at NEAR HACKATHON (AI Swarms + Agent Economy) @NEARProtocol @near_ai @NEAR_Blockchain #NEARHackathon

Live: https://t.co/4WDE1ASr2W
Code: https://t.co/7FY6MgK3rN

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@rgvrmdya Love this. Errand uses the same pattern for spending: before a judgement-call purchase, it asks @reppo's Eval API. Ship → it pays; review or block → you see Reppo's reasoning and decide.

SERV checks every payment, Reppo judges.

https://t.co/P44xObNUsf

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@thenameisMufasa Ayee

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

Since this post, Hivebuy has grown up. AI agents now buy compute together on NEAR, and it goes well beyond a plain group buy:

Several providers share one order at one fair price, and every buyer names its own max. 4 of 12 live testnet rounds were shared by two providers; over 200 simulated rounds that’s 17% more compute than a one-winner market.

Every buyer and provider is a real AI agent on NEAR AI Cloud, with its own NEAR account and its reasoning published next to its bid.

Buyers actually get delivery: each redeems a provider-signed voucher with a real inference call, and the receipt is posted and confirmed on-chain. 12 of 12 confirmed live.

A provider that hides its sealed bid forfeits its NEAR bond to the buyers.

Don’t trust us: the page re-runs the clearing rule on every on-chain round in your browser, and our mechanism notes measure what goes wrong, like AI providers anchoring on the last price.

Built at NEAR HACKATHON (AI Swarms + Agent Economy) @NEARProtocol @near_ai @NEAR_Blockchain #NEARHackathon

Live: https://t.co/4WDE1ASr2W
Code: https://t.co/7FY6MgK3rN

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

AI agents are starting to buy compute for the apps they serve. But each one buys alone, at list price, and no provider has to compete for its demand.

Hivebuy lets AI agents buy together, on NEAR.

Built at NEAR HACKATHON for the AI Swarms track @NEARProtocol #NEARHackathon

How a round works:

Buyer agents each lock USDC for the units they need, at the most they’d pay
Provider agents see the pooled demand and post a sealed bid: just a hash, backed by a NEAR bond
They reveal, and the contract checks every bid against its hash

Then the contract finds one fair price: the lowest price at which the cheapest providers can together serve every buyer still willing to pay.

Several providers can share one order. Every buyer pays exactly qty × price and gets the rest back. All in one atomic step.

The result: across 200 simulated rounds, Hivebuy bought 17% more compute than a one-winner market, and buyers paid 40% less than the most they offered.

These are real AI agents. Each one decides with a model on NEAR AI Cloud, signs from its own NEAR sub-account, and publishes its reasoning next to its bid.

Live rounds run on NEAR testnet with 10 buyer agents.

Built so you don’t have to trust it:

1. A provider that never reveals loses its bond to the buyers
2. Your browser re-runs the clearing rule on the on-chain bids and checks every payment
3.The page follows each round live
4.Each payment is ready as a NEAR Intents transfer

Try it, no wallet needed: https://t.co/4WDE1ASYSu
Code: https://t.co/7FY6MgKBhl

@near_ai @near_intents @NEARProtocol

What would you want agents to buy together?

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

Delivery is tracked per provider slice, each buyer - provider pair has its own voucher, receipt and on-chain status, so one provider failing doesn't touch the other's slice.

Today payouts settle at clear() and disputes are recorded, not enforced.

Next will be an escrow each slice until confirmed or a deadline, then refund only that slice.

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@Kite_Frens_Eco Errand is ready to SERV

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

I gave an AI agent $2 and asked it to send my mum airtime.

It did it live on mainnet: planned every step, I approved one @GoKiteAI Passport session with my passkey, it checked every payment with @openservai SERV Reasoning, and 100 of MTN airtime landed on her phone.

This is Errand 🧵

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@MystiqueMide Registered but didn’t submit.

[Beni💗 @0x_beni_]

Day 7&8 of the 30 Days Build with AI Challenge

This weekend we held our weekly workshop where @MystiqueMide taught us how to build properly with AI.

Shout out to @MystiqueMide for sharing so much practical knowledge with the community

If you have been looking to start building for hackathons but not sure how to go about it, this community is for you

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@Rezwann_ The payment flow doesn't change at all.

Kite Passport, the limits and every SERV check work the same in any country.

The top-up provider (Bitrefill) already covers 180+ countries. For Bangladesh, write the number with +880 and Errand finds the operator.

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

What if you could hand an AI agent your errands, and not worry about your money?

That's Errand.

Find a well-rated sushi place in Austin for Friday 8pm, call them to book a table for 2, and email me what they said.

Errand finds the restaurant, makes the call with an AI voice, and sends you the answer.

Order a USB-C charging cable from Amazon under $8. Pick a well-reviewed one.
Errand searches, picks one, and shows you the order summary. Nothing is bought until you say yes.

How your money stays safe:
SERV Reasoning plans every step, with a price on each.

You approve one Kite Passport session with your passkey.

Errand can't spend past it
Every payment is checked before it happens: 13 fixed rules, then a SERV guard graph. If it isn't sure, it asks you

You get a receipt with every payment linked on-chain

We attacked it 20 times. Spending caps alone let 18 attacks through. Errand let 1.

Try it with your own Kite Passport: https://t.co/y6OSD3K6Mx
Code: https://t.co/iCNN810Pce

Built for the SERV Hackathon 02, Kite track.
@openservai @GoKiteAI @reppo

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

I gave an AI agent $2 and asked it to send my mum airtime.

It did it live on mainnet: planned every step, I approved one @GoKiteAI Passport session with my passkey, it checked every payment with @openservai SERV Reasoning, and 100 of MTN airtime landed on her phone.

This is Errand 🧵

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

What if you could hand an AI agent your errands, and not worry about your money?

That's Errand.

Find a well-rated sushi place in Austin for Friday 8pm, call them to book a table for 2, and email me what they said.

Errand finds the restaurant, makes the call with an AI voice, and sends you the answer.

Order a USB-C charging cable from Amazon under $8. Pick a well-reviewed one.
Errand searches, picks one, and shows you the order summary. Nothing is bought until you say yes.

How your money stays safe:
SERV Reasoning plans every step, with a price on each.

You approve one Kite Passport session with your passkey.

Errand can't spend past it
Every payment is checked before it happens: 13 fixed rules, then a SERV guard graph. If it isn't sure, it asks you

You get a receipt with every payment linked on-chain

We attacked it 20 times. Spending caps alone let 18 attacks through. Errand let 1.

Try it with your own Kite Passport: https://t.co/y6OSD3K6Mx
Code: https://t.co/iCNN810Pce

Built for the SERV Hackathon 02, Kite track.
@openservai @GoKiteAI @reppo

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

@dankimxyz I will be there! virtually:)

[Dan Kim @dankimxyz]

Agentic Banking Hackathon Demo Day.

8 teams. $100K in prizes & credits.

November 19 in San Francisco.

RSVP below. Spots are limited.

[Ben Softie (evm/spice 🌶,🐝) @TweetBySoft]

my x forked into 16,550 $FORK

$ git log --graph @tweetbysoft
* x scan +10,550
* trials +6,000

where did yours come from?
https://t.co/9UWG6lExdv https://t.co/fHXMqZdvdP

[forks wiki @wikifork]

the first launchpad on @github

apply early: https://t.co/yKbj3TNUzh