Pinned Tweet
For the Agent-Native hackathon on @builderio , I built and Submitted Stigi ;The feedback board for product teams
Drop in hundreds of customer comments and an agent sorts them into themes. Drag two notes to correct it, and it re-sorts the rest to match.
Product managers and researchers struggle to turn hundreds of raw customer comments into themes. One-shot AI summaries fail because you can't correct them: fix one grouping by hand and the rest of the pile stays wrong.
Stigi lets you correct the agent's sorting by dragging two notes together. The agent names the theme, pulls in every matching note from across the board, and writes a one-line reason on each note it moved.
How it works
1. Load 200 sample comments (or import a CSV). The sample board starts imperfect: six clusters and 50 notes unsorted.
2. Drag two related notes into a new cluster, for example "which plan am I even on?" and "the billing page makes no sense".
3. The agent names the cluster and moves the matching notes in. In my recorded run it moved 14, each with an "Agent:" reason chip.
4. Click Sort with agent to file the rest in batches of 25, then Report for themes, quotes and a next step you can edit and export as Markdown.
- Corrections that stick: one drag teaches a theme, with no prompt to write.
- A reason on every move: the agent can't move a note without a one-line reason.
- An insight report: themes by size, verbatim quotes and a suggested next step.
- Safe to try: sign-in required, one private board per account, 40 agent turns a day.
How it uses Agent-Native:
the board UI and the agent call the same defineAction() actions (one move-notes for a drag and for an agent move, which must give a reason), share your selection through app-state and view-screen, and the agent opens the report with navigate.
Pre-existing work and what changed:
built on the Agent-Native "chat" template and toolkit (@agent-native/core 0.226.0) plus open libraries; everything Stigi-specific (actions, schema, board UI, agent skills, turn cap, landing page, logo, deployment) was built during the event on 2026-11-09.
App: stigi.suitz.xyz
Code: github.com/Dotman-Bei/stigi
Slide: stigi-slides.vercel.app/
Feedback welcome, especially on whether dragging two notes feels like enough to teach the agent a theme on your own data.
Below is the 1:38 minutes demo walkthrough @Steve8708
Heisbei retweeted
In 2018, Arkansas cut about 18,000 people from Medicaid. Most of them qualified. They lost their coverage to paperwork.
On December 31, 2026, a rule like that goes national: 80 hours of work, school or volunteering, or $580 of pay, every month, with proof.
So I built Eighty. It does that paperwork from the phone people already use.
Send a pay stub photo or a voice note in English or Spanish on Telegram. Eighty reads it and proposes a line. Your month fills in on the web as two rings. Tap "Do my month" and it asks you the one question only you can answer, then gives you a proof packet, cited to the law, ready to send to your state.
You can even connect Claude and ask what's left in your month.
The agents propose. Only you confirm. Nothing counts until you do.
Why it only works on Agent-Native: Eighty has four ways in. Your buttons on the web, the helper on the web, Telegram, and Claude. All four run the same 27 actions, on one database, with the same check of who is asking. Built any other way, that's four separate systems that could disagree about whether someone keeps their health coverage.
And the rule that matters most is set on the action itself: the agents propose, only you confirm. No agent, in the app, in Telegram or in Claude, is ever given the confirm button.
Built on Agent-Native for the Builder.io Agent-Native Hackathon.
Try it, no sign-in needed: tryeighty.space
Telegram: @TryEightybot
Code: github.com/Jennycruzy/eighty
Heisbei retweeted
𝐏𝐀𝐕𝐈𝐋𝐈𝐎𝐍 is a floor planner for events where event organizers and Ai agents collaborate to design, optimize and manage venue layout in real time.
It allows users to create scaled floor plans, arrange seating , stages and equioment and explore layouts in 2D or 3D. It’s intelligent layout engine automatically checks spatial constraints , seating capacity, safety requirements and costs ensuring every design is pratical and ready for execution.
Rather than making changes without permission, the ai proposes improvements that users can review, compare and approve, keeping humans in control throughout the planning process.
𝐇𝐎𝐖 𝐈𝐓 𝐔𝐒𝐄𝐒 𝐀𝐆𝐄𝐍𝐓-𝐍𝐀𝐓𝐈𝐕𝐄
- One action layer: the canvas and the agent call the same defineAction() contracts (move-objects, add-objects, propose-layout, export-setup-pack and others). There is no agent-only path.
- Application state: navigation and the current selection are synced, and view-screen returns them verified, so "move this closer to the stage" means the chair you selected.
- A human gate in the contract: apply-proposal and restore-version are UI-only actions, hidden from the agent's tools and MCP. The agent proposes; a person applies.
- Every write is versioned and audited in PostgreSQL through the framework's database conventions. Manual editing works with no model key.
- Each visitor gets an isolated guest sandbox with rate limits, so the live demo needs no sign-up.
Pavilion : pavilion-nine.vercel.app
GitHub Repo: github.com/Enoch208/pavilion
𝑯𝒆𝒓𝒆 𝒊𝒔 𝒂 2 𝒎𝒊𝒏𝒖𝒕𝒆 𝒗𝒊𝒅𝒆𝒐 𝒘𝒂𝒍𝒌𝒕𝒉𝒓𝒐𝒖𝒈𝒉 𝒐𝒇 𝒉𝒐𝒘 𝒕𝒉𝒆 𝒑𝒓𝒐𝒅𝒖𝒄𝒕 𝒘𝒐𝒓𝒌𝒔
@Steve8708 @builderio @tempoimmaterial
Heisbei retweeted
For the Agent-Native hackathon on @builderio , I built and Submitted Stigi ;The feedback board for product teams
Drop in hundreds of customer comments and an agent sorts them into themes. Drag two notes to correct it, and it re-sorts the rest to match.
Product managers and researchers struggle to turn hundreds of raw customer comments into themes. One-shot AI summaries fail because you can't correct them: fix one grouping by hand and the rest of the pile stays wrong.
Stigi lets you correct the agent's sorting by dragging two notes together. The agent names the theme, pulls in every matching note from across the board, and writes a one-line reason on each note it moved.
How it works
1. Load 200 sample comments (or import a CSV). The sample board starts imperfect: six clusters and 50 notes unsorted.
2. Drag two related notes into a new cluster, for example "which plan am I even on?" and "the billing page makes no sense".
3. The agent names the cluster and moves the matching notes in. In my recorded run it moved 14, each with an "Agent:" reason chip.
4. Click Sort with agent to file the rest in batches of 25, then Report for themes, quotes and a next step you can edit and export as Markdown.
- Corrections that stick: one drag teaches a theme, with no prompt to write.
- A reason on every move: the agent can't move a note without a one-line reason.
- An insight report: themes by size, verbatim quotes and a suggested next step.
- Safe to try: sign-in required, one private board per account, 40 agent turns a day.
How it uses Agent-Native:
the board UI and the agent call the same defineAction() actions (one move-notes for a drag and for an agent move, which must give a reason), share your selection through app-state and view-screen, and the agent opens the report with navigate.
Pre-existing work and what changed:
built on the Agent-Native "chat" template and toolkit (@agent-native/core 0.226.0) plus open libraries; everything Stigi-specific (actions, schema, board UI, agent skills, turn cap, landing page, logo, deployment) was built during the event on 2026-11-09.
App: stigi.suitz.xyz
Code: github.com/Dotman-Bei/stigi
Slide: stigi-slides.vercel.app/
Feedback welcome, especially on whether dragging two notes feels like enough to teach the agent a theme on your own data.
Below is the 1:38 minutes demo walkthrough @Steve8708
Heisbei retweeted
As AI makes it easier to find and exploit vulnerabilities, security is becoming a bigger concern for every tech company.
Lucas, Jito's co-founder and CEO, compares the growing wave of security threats to the early days of COVID, when warning signs kept appearing before things escalated.
He believes we're heading into a period where security needs to be every company's top priority, even if that means slowing down to fix what's already been built.
“It's gonna get way worse before it gets better.”
Heisbei retweeted
Every API crashes on the inputs you never test for: a negative number, an array where a string belongs, an injection string and those unhandled 500s leak stack traces straight to your callers.
Happy-path test suites never catch them, because they only send the inputs you already expected.
And even when you do find a crash, the slow part isn't finding it, it's diagnosing the real exception and writing the fix.
That diagnose-and-fix step is exactly where an AI agent earns its place, so I built on @builderio's Agent-Native framework, where every capability is a typed action the agent and the UI share; meaning the agent can't fake a result, it can only run the same real code you do.
What I built is ReqLab: an API, webhook, and chaos-testing studio where the agent breaks your APIs on purpose
you point it at an endpoint and it runs a Chaos Matrix of adversarial payloads:
- injection,
- type confusion,
- negatives,
- oversized strings,
- bad auth - until something throws a real 500 with its actual stack trace.
then the agent takes over: it reads the trace, writes a Zod contract guard, drops it in as a gateway, and re-runs until every red 500 becomes a clean 422.
Hit "Bug Hunt" and it attacks in escalating rounds on its own, sharpening each batch at whatever leaked a trace.
It even imports a live OpenAPI spec and found a genuine unhandled 500 in the public Swagger Petstore; an API I didn't write
the guard compiler never evals your schema, external calls are SSRF-guarded, and one click exports the guard as Express/Hono/Fetch middleware for the real service.
The whole find-the-crash, prove-the-fix loop runs from a single chat message, live.
And it's not just agent-driven - it's agent-reachable: ReqLab exposes a secured MCP server, so you can drive the whole thing from Claude, Cursor, or ChatGPT, not only its own chat.
Live app: usereqlab.vercel.app
MCP endpoint: usereqlab.vercel.app/mcp
#AgentNative #BuildWithAI
. @Steve8708 @tempoimmaterial
Heisbei retweeted
An AI agent that earns money can't own it, because someone always holds the key.
Incorpr fixes that on @NEARProtocol: every AI company is an NFT with its own vault account, and that vault has zero access keys.
Through Chain Signatures, the vault controls real Bitcoin, Solana and EVM addresses, and it has already signed a Bitcoin spend on mainnet.
Customers pay it per call in USDC, and each answer comes signed by a NEAR AI Cloud TEE.
Its treasury agent rebalances through NEAR Intents, but the contract decides what can leave.
We let anyone try to talk the agent into robbing its own company: 11 attempts so far, 11 refused on chain.
The owner can sell the entire company, or borrow USDC against it, in a single transaction.
It's live on mainnet now, with paid calls, sales and loans, and a listing on NEAR's Agent Market.
All 75 transactions are public and re-verifiable from the repo.
Try it: tryincorpr.vercel.app
. #NEARHackathon
Heisbei retweeted
most prediction agents forget. Every session they pay for the same tips and trust every source the same, so the sources that keep being wrong keep getting paid.
I built Calibro for #TotalRecall, the @tatum_io × @WalrusProtocol hackathon. it keeps score of every source on Walrus Memory and pays only the ones that earned it.
Each signal answers HTTP 402. Calibro pays on @SuiNetwork mainnet, and the server checks every payment through Tatum's Sui gateway before it releases the signal.
when Polymarket resolves a game, a scheduled job scores every signal it bought and writes the post-mortem back to Walrus. Nobody presses a button.
the results are real: 41 games settled, 144 signals paid, and 29 payments refused from memory. One source scored worse than a coin flip, so Calibro stopped paying it.
People can make picks too. They sign them with their wallet, and the picks are graded when the game ends. The best pickers become a paid source.
it hasn't beaten Polymarket yet (Brier 0.194 vs 0.185), and I say so. what it has learned is who is worth paying.
article: medium.com/@mrnetwork/teachi…
🔗 trycalibro.xyz
Heisbei retweeted
AI agents are starting to hold money, and today that usually means handing the agent a key and hoping the prompt holds. One prompt injection or bad tool call and the funds are gone.
I built Leash at NEAR HACKATHON: the agent never holds the money. It can only ask a NEAR smart contract to pay, and the contract decides.
What NEAR does here: the Rust contract is the vault and enforces the rules onchain. The agent has its own NEAR account and can only call request_payment. The owner signs in with a NEAR wallet to approve, edit the rules or pause.
Rules the contract enforces:
• max per payment and per day
• recipient allowlist
• risk score check
• owner approval above a set amount
• emergency pause
• every attempt, even blocked ones, logged onchain
In the video:
→ 0.1 NEAR to the vendor: within rules, paid
→ 0.3 NEAR: held until the owner approves, then paid
→ unknown wallet: blocked
The AI can be fooled. The contract can’t.
Try it yourself (testnet): web-mu-eight-75.vercel.app/
Code: github.com/BeautyBenedict/le…
What would you want an AI agent to never be able to do with your wallet?
Built at NEAR HACKATHON
@NEARProtocol #NEARHackathon
Heisbei retweeted
24 hours left to submit your Vibeathon 02 build.
Submissions close Mon 9 Nov at 22:00 UTC. Send us the link to your live app, not a demo video.
Please remember, If you started from an existing app, tell the judges what you built during the week.
app.vibestarter.xyz/vibeatho…
Heisbei retweeted
Introducing Joust, my entry for the Hermes Agent Plugin Hackathon @NousResearch
Ask three AI agents to fix the same bug in one folder and they overwrite each other.
Joust gives every contender its own Git worktree and its own Hermes subagent, and it can run each one on a different model.
In the demo, Claude Sonnet 4.6, GPT-5.5 and Gemini 3.1 Pro raced to fix a real order-book bug.
All three passed the official tests, three runs each, in under two and a half minutes.
Then Crossfire fired each model's own tests at its rivals.
Gemini had outscored GPT on points, but both rivals caught it dropping leftover quantity on partial fills, a bug the official suite never tested. Gemini fell to last.
Claude won with 91.5 points, and its fix ran in 3.65 ms against 260 ms for the old code.
The winner merges only after you approve it in Hermes's own approval card.
Every attempt is kept as a Git branch, and one call rolls the merge back.
Real run, nothing mocked.
Built on Hermes hooks, middleware and the subagent lifecycle, with a full arena inside Hermes Desktop.
May the best patch win.
Code: github.com/mrnetwork0001/Jou…
Catalog PR: github.com/NousResearch/herm…
Picture a token community. The chat is full of "someone should build that." Weeks later, nothing exists. And if you do hand the job to an AI agent, you can't check what it did, what it touched, or what it cost.
I built Night Crew for @vibestarterxyz 02. $VIBES holders vote on what gets built, an AI crew builds the winner overnight, and every build gets a public receipt.
HOW IT WORKS
1. Open the site and connect a wallet. Holding $VIBES gives your vote weight. A wallet with none still counts as 1.
2. Sign a message to vote on a proposal or to propose your own. It costs no gas, and you can change your vote until the close.
3. At 22:00 UTC a GitHub Actions job closes the vote, runs the crew, verifies the result and publishes it.
4. The winner appears as a new floor in the tower, with a receipt attested on Base through EAS.
The agent has no authority. It has three tools: write a file, read a file, and run a floor check. It has no shell. A static gate checks its output in-process, and the crew job never executes anything the agent wrote.
Each floor is served as its own sandboxed page, under a hashed CSP, with no access to your wallet. A veto address can pull a bad one.
The receipt records the shift, the outcome, the model, the cost, and hashes of the transcript and the floor. Anyone can check it on EAS.
- 5 floors are lit.
- Shift 5 built a quiz: 6 of 6 tests passed, $0.07, attested.
- Each shift has cost between $0.06 and $0.10.
- 6 real wallets have voted or proposed, with 6 votes signed.
Live: nightcrew.suitz.xyz
Code: github.com/Dotman-Bei/nightc…
Below is the Demo wallthrough (94 s) @CyphrGM @ProlabCH
Heisbei retweeted
What if your USDT could earn interest through Nimiq Pay without having to leave the ecosystem?
That's the idea behind CURRENT, a @nimiq Pay Mini App I built for the @miniappscomp Competition (Earn Track).
Current connects Nimiq Pay users to Aave V3 on Polygon, allowing them to supply and withdraw USDT while accessing variable lending interest rates.
With Current, users can:
- Check current USDT supply rates.
- Supply and withdraw USDT.
- Track their lending positions.
- Review and authorise transactions directly through their wallet.
The goal is to make lending more accessible to Nimiq users and bring potential interest-earning opportunities into the Nimiq ecosystem
Try it: current.oladapodamiey.worker…
Watch : youtu.be/EDPMHaYVrj0
Code: github.com/demonchant/curren…
Heisbei retweeted
Handing over a blockchain project should mean handing over the control that makes it run.
You can give someone the code, the website and all the project files, but that does not automatically give them permission to manage the contracts. They might have everything except the ability to actually run the project.
Introducing Writ.
I built Writ at NEAR HACKATHON to turn supported project permissions into transferable digital keys.
Those keys are NFTs. But instead of only displaying a picture, they let their holders do something.
Here’s a simple example.
Ada manages an NFT collection. She creates a Writ and gives the account behind it permission to administer that collection.
While Ada holds the Writ NFT, she can use its approved commands to issue collectibles.
When Ben takes over, Ada sends him the NFT.
Ben can now use those commands. Writ rejects Ada’s former-holder access, and temporary permissions granted under her ownership are cancelled.
The collection stays where it is. The NFT changes hands, and control through Writ follows.
No sale is required. A handover can simply be a transfer.
And this goes beyond NFT collections. Writ currently supports five profiles:
• Managing a compatible service contract
• Exercising a DAO council seat
• Administering a compatible NFT collection
• Issuing tokens through a compatible token contract
• Managing payouts from a native-NEAR treasury
You can also lend someone a specific action with fixed limits and an expiry, without handing over the entire role.
Before accepting a Writ, someone can inspect its holder, approved actions, temporary permissions, available funds and recorded history.
Selling is optional. When a sale is used, a buyer can require declared conditions to pass before the purchase completes.
Developers can also use Writ from their own applications through its standalone TypeScript SDK. The holder does not need to stay on Writ’s website to exercise the permission.
The technical part:
Each Writ has a dedicated NEAR account with no access keys. NEAR smart contracts check the NFT’s current holder before forwarding an approved command to the external application.
Transferring the NFT updates that authorization and invalidates the previous ownership’s temporary permissions. Supported ownership roles also allow a verified hand-back into the holder’s own wallet.
These are reviewed integrations with specific contract interfaces, rather than a promise to control every contract automatically.
The attached demo shows the full journey using real transactions on NEAR Testnet: create, grant permission, use, lend, transfer, execute from another application and hand control back.
The client supports both testnet and mainnet. Mainnet users can connect a compatible Writ deployment on their chosen network.
Try Writ: writ-app.vercel.app/
Explore the code: github.com/ELLA0VICTOR/writ
What permission would you want to use, lend or hand over this way?
@NEARProtocol #NEARHackathon
Heisbei retweeted
For the Flip the Switch Hackathon by @WeMakeDevs
and @switch_ai_ , I built and submitted Invoice airlock:
An accounts-payable desk on Switch that stays safe when the agent reading vendor invoices is hijacked: outside content stays in a room with no reach, only validated typed fields cross to the room that can pay, and only the controller can release money.
An AP desk on Switch that assumes the agent reading invoices is already hijacked. 9 attacks replayed, the ledger never moved.
The point is that Switch's rooms and owners hold when the agent is fooled.
For the full write-up, visit: medium.com/@bamigboyeemmanue…
Live app: invoice-airlock-site.vercel.…
Code: github.com/Dotman-Bei/invoic…
below is the 2:40 demo walkthrough.
#HowISwitch.
Heisbei retweeted
I built Vigil for the Flip the Switch Hackathon by @WeMakeDevs and @switch_ai_: an incident response workflow where Claude Code fixes production, and only a person can ship the fix.
When a release breaks checkout, a sentinel agent takes command in a Switch room and hands the failing request to Claude Code in our Mattermost.
Claude Code reproduces the bug and writes the fix, then a second agent checks it with the real tests and a replay on staging.
The release gate is plain code, not a model, and it refuses every agent, including the one that wrote the fix.
Only a person typing "approve" in chat can ship it.
On the live server, INC-2046 went from alert to closed in 77 seconds, and the agent's own approval was refused on the way.
On the next incident, my teammate Nzubechukwu approved the release from their own account.
Every screen in the demo is a recording of the live system, not a mock-up.
Try it: vigil.netlayerlabs.com
Code: github.com/NetLayerLabs/Vigi…
#HowISwitch
For the Flip the Switch Hackathon by @WeMakeDevs
and @switch_ai_ , I built and submitted Invoice airlock:
An accounts-payable desk on Switch that stays safe when the agent reading vendor invoices is hijacked: outside content stays in a room with no reach, only validated typed fields cross to the room that can pay, and only the controller can release money.
An AP desk on Switch that assumes the agent reading invoices is already hijacked. 9 attacks replayed, the ledger never moved.
The point is that Switch's rooms and owners hold when the agent is fooled.
For the full write-up, visit: medium.com/@bamigboyeemmanue…
Live app: invoice-airlock-site.vercel.…
Code: github.com/Dotman-Bei/invoic…
below is the 2:40 demo walkthrough.
#HowISwitch.