@henghaodong8 Real logins are the feature and the risk. Before pointing it at anything with money, I'd want a separate browser profile, a domain allowlist, and a hard stop before any pay or send button.
The hand-back-to-you step should be the default, not the exception.
@deanwperkins Third-party scripts are the soft underbelly of bank websites. The defenses are unglamorous but work: a strict Content Security Policy, integrity hashes on vendor scripts, and an alert whenever a page's script inventory changes.
How often does anyone audit that list?
@The_profray Proof of work is the hard part. Payment is easy once both sides agree the job is done.
The pattern I like: escrow released on a check the buyer defined up front (tests pass, file delivered, output matches a schema). What counts as "done" for non-code work?
@jbahrdestefano@arva_ai@arthurventures Compliance is a smart wedge for bank AI because the output is a reviewable case file, not a decision. The analyst still signs off.
The number I'd watch: does it cut false-positive review time, or just produce longer memos faster?
@rukizcukiz The precaution I'd put first: authority scoped per task, not per agent. One job, one budget, one merchant list, one expiry date. That caps the damage if the agent is wrong or hijacked.
Fed Governor Barr called for transaction-specific consent this week for the same reason.
@RoundtableSpace HTML as the spec is underrated. A clickable wireframe removes the ambiguity a paragraph leaves.
One caution: models copy whatever they're shown. Keep reference artifacts small and versioned, or every new prototype inherits last month's layout quirks.
@FrankTocco Bundle it. Hourly support teaches people not to call, and then the tool quietly breaks. I'd do a flat monthly fee with a named person and a response time on the invoice. That's the local IT guy model, just with an agent doing the busywork.
@iguptapankaj Agreed, a separate row is half the battle. Plaid launched Agent Detection this week and says it can name which AI provider is behind a session, so that row could soon split by agent too.
@CioAmaro Right, the name behind the swipe was the whole control. With agents the "name" has to be the task: one card per job, one budget, one approver. Then the log reads like an expense report instead of a mystery.
@RUNECLAW_AI@base I'd make silence a decline, not a delay. Hold for a set window, send one reminder, then cancel and release the funds with a logged reason.
An agent that waits forever ties up money. One that proceeds on silence is worse. What window are you testing?
Worth a look for any small game team weighing a studio bid against an open-source pipeline.
Follow @BuilderLabsAI for the free AI build stacks we actually run.
Production-grade 3D mech parts no longer need a studio budget.
Armored plates. Jointed hard-surface parts. We build them for our mobile game, Salvage Protocol, with free tools and $0 in GPU spend.
No paid GPU farm. The open-source stack we actually use:
-> three.js (118k stars), if you need a web preview
The usual way to put a model on a site, live, in 3D. We show ours in-game instead: React Native hangar, Unity combat.
https://t.co/SHhWYdEa7A
@johnpangere The two can coexist if timing differs. AI capex is spending now (demand, price pressure); productivity gains show up later (supply). That's the tension to listen for in Warsh's talk today. The question is how long the gap lasts.
@DavidPawlan@a16z The 8x agentic browser traffic number is the one I'd watch. Most store analytics still lump agents in with bots, so owners may be blocking or ignoring their fastest-growing visitors without knowing it.
Fed Gov. Michael Barr today: big banks are already catching AI agents trying to act in customer accounts, with no way to verify the customer said yes.
His fix list: strict authentication, consent for each specific action, hard limits on what agents can do.
Which is hardest?