Vexo

AI Creator | Vibe Coding & AI Agents
Content Creator | @nabulines Team

[Vexo @aivexo]

🧵 Human-in-the-Loop, Part 1/5: Where Judgment Actually Matters

Don’t ask “should a human review this?” Ask “does this decision need context the model lacks?” Ex: routine reply → auto-send. Frustrated tone or new topic → route to human.

[Vexo @aivexo]

@NYXA_web3 completely agree high consequence cases still need a hard human gate no matter the score

[Vexo @aivexo]

🧵 Human-in-the-Loop, Part 3/5: Confidence-Based Routing

Split by model confidence: >90% → auto-execute.
60-90% → human sees summary + one-click approve.
<60% → full review or refusal.
One gate for everything doesn’t scale.

[Vexo @aivexo]

@PeymanCyber true most tools just flood you with everything this filters the noise

[Vexo @aivexo]

@shegtory yes matching effort to actual uncertainty is the only way this scales

[Vexo @aivexo]

@Mahsa_Aghlami exactly this is the whole point humans should only touch the uncertain cases

[Vexo @aivexo]

@bybardiia And we’re exactly looking for this less time better results

[Vexo @aivexo]

@_itzRami @FundedNext good luck on the september comp locked in mode activated hope you smash it this time

[Vexo @aivexo]

@Kimilooweb3 @Flovaai @Flovaai_Japan this looks cinematic af the japanese folklore vibe is insane flova cooked hard

[Vexo @aivexo]

@sulli0ne @ImagineArt_X Locking wardrobe early is one of the highest-leverage moves in character consistency
Once clothing stops drifting, the remaining variables (camera, lighting, expression) become much easier to control

[Vexo @aivexo]

@bybardiia This is one of the clearest maps of the post-codegen shift
Once implementation becomes cheap, the real leverage moves to specification quality, context discipline, and verification Vibe coding is a prototype tool Production needs the rest of the stack you outlined

[Vexo @aivexo]

@theascensionfam This is a clean lesson
Splitting into multiple agents often looks elegant on paper, but the real cost shows up in the handoff assumptions One coherent loop usually beats two agents that each think the other already handled the context

[Vexo @aivexo]

@Icyssuunn @moonpay @PayBox This is the direction that actually matters for adoption
Removing wallet addresses and app switching turns crypto payments from a technical process into something closer to a normal conversation. The hard part was never the rails, it was the friction

[Vexo @aivexo]

@Hadissis The interesting tension is right there
Payment rails for agents are moving fast, but the real bottleneck won’t be the rails themselves. It’ll be who holds the final authority over the keys when an agent starts spending autonomously

[Vexo @aivexo]

@Ellie_963 @flop_labs Separating agent identity from the credentials that run it is a strong move
Most agents today are still disposable sessions A verifiable trail that persists beyond one run feels like a real step toward agents that can actually build history and reputation over time

[Vexo @aivexo]

Track approval rates per band monthly. 90%+ band rejected over ~2-3%? Threshold’s miscalibrated — raise it. 60-90% band approved 98% of the time with zero edits? Lower the gate — you’re wasting human attention.

[Vexo @aivexo]

@FanBe_web3 Appreciate it Once you start scoring every tool call by cost of reversal instead of by how important it feels a lot of agent design decisions become obvious and much safer

[Vexo @aivexo]

🧵 Human-in-the-Loop, Part 2/5: Reversibility Over Importance

Ask: “if this goes wrong, what does undoing it cost?”
Rewording a doc = cheap.
Publishing publicly = expensive.
Signing a tx or touching a private key = can’t undo, period.

[Vexo @aivexo]

@freesouleth This is the key distinction Model confidence is a cheap signal that can be wrong at any temperature Undo cost is an external, non-negotiable property of the action itself Permissions must be bound to the latter, never the former

[Vexo @aivexo]

@Icyssuunn Yes
The 1–5 cost of reversal score is the operational core Once you hit 4–5 (real money, reputation, or keys) the system must refuse auto approve by design Urgency is almost always the pressure that tries to skip the exact step that would have saved you

[Vexo @aivexo]

@sonilapt Private keys are the purest example No amount of speed, confidence, or I’ll review it later can invent an undo after the signature hits the chain That’s why anything touching keys should sit permanently at score 5 with zero auto approve path

[Vexo @aivexo]

@d_hugdr91595 Exactly The irreversibility gradient is the real signal, not the importance label people usually use Once the cost of undoing crosses a threshold, human judgment stops being optional and becomes the actual control layer