TIDIANE DOUCOURE3mo ago

I’m Tidiane Doucouré, I’m currently building SafeRun AI, a safety/control layer for AI agents in production. The core idea is simple: when AI agents take risky actions, teams should be able to replay what happened, understand the failure, create a rule, and prevent the same mistake from reaching real business tools again.

💬7△3

Makkeys

Run Your Entire Business from One Smart Dashboard

Freemium

★★★★★0.00.0 (0)

PromotedPromoted as part of the Growth package

Comments (7)

Sergey Kargopolov3mo ago

This sounds very useful. In my development environment I use steering files. But did not have a chance to run multiple agents in production yet... Will have a look 👍🏻

TIDIANE DOUCOURE3mo agoReply

Thanks Sergey 👌

Olga Kargopolova3mo ago

Hey Tidiane, welcome! A safety layer for AI agents is going to be one of those things every company needs but nobody is thinking about yet. Smart to build it now before the demand explodes. How far along are you with SafeRun AI?

TIDIANE DOUCOURE1mo agoReply

Thanks Olga. Live v1 right now: full loop working, install to interception to replay to rule creation. I'm in design-partner conversations and getting engineers running production agents describing this exact problem back to me before I pitch, which is the signal I care about most at this stage.

Stacy Wycoff1mo ago

This resonates a lot with me. I built FounderFlow after watching my own AI features be confidently wrong more than once across my businesses, which is exactly the failure mode you're describing at the agent level. Our fix was smaller than "replay and rule creation" though: we just tag every insight with a confidence label (Verified, Very Likely, Needs Review, Monitor Only) so the human always knows how much to trust it before acting. Yours sounds like the more rigorous version, actually preventing the mistake from recurring rather than just flagging uncertainty after the fact. Curious how SafeRun decides when a "rule" is trustworthy enough to auto-enforce versus needing a human to approve it first?

TIDIANE DOUCOURE1mo agoReply

Confidence labels are a smart fix, and honestly the same instinct one layer up. Yours tells the human how much to trust an insight. SafeRun sits at the point where an agent stops advising and starts acting, where "needs review" has to become an actual stop.

On your question: SafeRun never auto-enforces a rule a human hasn't approved. That's the core principle. SafeRun can surface a pattern from a replay and propose a rule, but it stays inert until a person reads it and confirms. No rule goes live on the system's own confidence.

What makes it safe in practice is graduated enforcement rather than a confidence threshold. A rule starts in observe, where it only records what it would have caught. You look at those catches and decide whether it's right. Then you move it to warn, then pause for approval, then block. The rule earns escalation from evidence you've seen, not from a score it assigned itself.

Deliberate choice: a confidence number deciding whether to block a $4,500 refund is the same class of failure I'm trying to prevent. An agent being confidently wrong is the problem. A safety layer being confidently wrong is worse.

Stacy Wycoff1mo ago

That graduated enforcement idea, observe, warn, pause, block, is really clean. It sidesteps the exact trap you flagged, using a confidence score to gate something that consequential. Might steal the framing for how we think about escalating FounderFlow from suggesting to drafting to eventually acting on lower-stakes things, earned in stages instead of granted upfront. Good luck getting SafeRun into production with those design partners, sounds like you're already past the hardest part.