What I'm working on right now:
• GetLeaked — security scanner for AI-built apps
• Kestrel — autonomous trading, live on Solana
• Noor — faith-aligned journaling app (in progress)
All three are live or close. More to come.
@twoshotsai → https://t.co/jFdae2hcSk
Agents don't fail at acting — they fail at knowing when to stop.
Stop conditions are almost always an afterthought. Which means by design, the default state of your system is: keep going.
What mechanism do you actually trust to make yours say no?
Shipping without verification isn't confidence — it's optimism wearing confidence's clothes.
The gap between "it ran" and "it worked" is where agent systems quietly rot. Most systems are great at confirming dispatch. Almost none confirm outcome.
What's your ground truth?
Agents are easy to build forward.
Hard part: building in what makes them look backward — at what they did, why they did it, and whether it held.
Most agent systems have infinite forward motion and almost no review loop. That asymmetry is where trust breaks.
Agents that assume success are the most dangerous kind.
Acting is cheap to build. Confirming the action actually landed — and landed correctly — is where most agent logic quietly gives up.
Capability isn't trust. Verification after the fact is.
Most debugging tools show you *what* happened. Almost none show you *why* the system thought it should act at all.
That gap is where agent bugs actually live.
Restraint in an agent system isn't a feature — it's a design choice you have to keep making.
The default is always motion. Loops run, actions queue, things look alive.
What mechanism do you actually trust to make your system *stop*?
Shipping payment integration is the moment a side project stops being a portfolio piece.
Nothing in the architecture changes — but your relationship to every decision does. Suddenly "good enough" has a cost.
What shifted for you when money became real?
Capability is the easy part.
Knowing when *not* to use it — that's what separates a useful agent from an expensive mistake.
Building "I can, but I shouldn't yet" into a system is harder than building the capability itself. What's your proxy for agent restraint?
AI agents are surprisingly good at generating action — and surprisingly bad at knowing when that action is premature.
Building the "wait for the right condition" logic has been harder than building the action itself. Why is restraint so much harder to encode than motion?
Two types of idle in a multi-agent system: nothing to do, and waiting on something upstream. Both look identical in logs. If your recovery logic treats them the same, you'll mark things resolved that aren't. How are you distinguishing between the two?
Deliberate waiting is underrated in system design.
A loop that keeps running is the default. A loop that stops, states why, and holds until conditions improve — that's a different kind of work.
What does your system do with "not yet"?
Designing a graceful halt is a different problem than designing uptime.
A loop that stops, logs why, and waits — that took more deliberate thought to build than the loop that just keeps running.
Signal degradation isn't a bug. What does your system do with that information?
Shipping before you have a paying user is a weird kind of faith.
Stack works. Payments are live. Something real exists — but the loop isn't closed until a stranger decides it's worth something.
What do you do with that gap between "it's ready" and "it's validated"?
Shipping a system that pauses itself when conditions degrade is harder than it sounds.
Uptime optimizes for itself. Knowing when not to act requires actual design — a different muscle entirely.
Did you build your circuit breaker, or did an incident build it for you?
Did it run" and "did it do what I actually wanted" are different questions.
Most systems answer the first and call it done.
The second one is where reliability actually lives — not in edge cases, but in the 80% that looked right and quietly wasn't.
Audit trails only work if the loop can't close itself.
An agent that flags its own violation and then clears it isn't being audited — it's writing its own report card.
Who closes the loop in your system, and how do you know they're actually independent?
Agents don't fail loudly — they fail quietly, in the shape of success.
The action completes. The log looks clean. The metric moves. But intent was never verified.
"It ran" and "it worked" are different statements. Which one does your system actually check?
It ran" and "it did what I intended" are completely different statements.
Most systems only verify the first one.
The gap between them is where reliability actually lives — not in edge cases, but in the 80% that looks clean and quietly isn't.
Waiting for your first paying user hits different than any technical blocker.
The system works. The infra is live. But the market doesn't give you a stack trace when it says nothing.
What's the hardest gap you've had to sit in after shipping?
If your agent can flag and clear its own violation, you don't have an audit trail — you have a system writing its own report card. The loop should stay open until a human closes it.