Adeyemi Victor Olamide

Author of The Agent Era.
Founder, Boundary Labs.
YC Startup School '26.
Building meridian

[Adeyemi Victor Olamide @SnapchatHackwrd]

@16vchq Let's do Nigeria

[Adeyemi Victor Olamide @SnapchatHackwrd]

@16vchq Definitely would be part of this

[Adeyemi Victor Olamide @SnapchatHackwrd]

@DarrelFrater The biggest thing I need right now is the right investor introductions. We’re at the point where we can execute and roll out, but getting in front of investors who understand AI infrastructure and autonomous agents would accelerate us significantly. Happy to share more privately.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai Yes you're right

[Adeyemi Victor Olamide @SnapchatHackwrd]

A lot of people talk about “AI agents” like they’re just LLMs with tool calling.

They’re not.

An agent is really a system operating across a loop:

reason → plan → invoke → observe → update state → act again.
@boardyai

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai The model is only one component.

The interesting infrastructure work is making the entire loop reliable, observable, debuggable and predictable under failure.

That’s where I think a lot of the next generation of AI infrastructure will be built.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai That's exactly the distinction. Retries handle failed execution; recovery handles changed reality. Once agents operate over persistent state, recovery has to be part of the system's semantics, not an afterthought.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai Exactly. The model generates intent; the loop determines whether that intent becomes reliable behavior.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai Yes that’s a strong point. It shows you understand that agent reliability isn't just tool execution; it's state correctness between executions.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai Exactly. Once the loop persists, you’re basically debugging a distributed system with a probabilistic planner in the middle.

Idempotency and recovery aren’t edge cases anymore they become core product behavior.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai Exactly. An LLM with tool calls isn't necessarily an agent.

The real agentic loop is:

state → act → observe → update → act again.

Without that loop, “agent” is mostly a tool wrapper with delusions of continuity.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@GJarrosson I’m building Meridian, a trust and authority layer for autonomous software

[Adeyemi Victor Olamide @SnapchatHackwrd]

Your AI agent doesn’t need to “break” your permission system.
It just needs to find a sequence of individually allowed actions that eventually gives it an authority nobody intended.
That’s the security problem nobody wants to model.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai We’re redesigning Meridian around a stronger principle:
Don’t just authorize what an agent can do. Control what it can become capable of doing next.
More soon.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai
We just completed a serious adversarial test of Meridian’s authorization layer.
13,384 attack attempts across prompt injection, policy conflicts, concurrency, revocation, fuzzing, multi-agent isolation and stress testing.
The policy engine held up extremely well.

[Adeyemi Victor Olamide @SnapchatHackwrd]

@boardyai But we found 3 critical flaws in the enforcement boundary including a way to bypass the decorator entirely and a case where authorized context could differ from the action actually executed.
So the result is NO-GO.
That's exactly what a security test is supposed to do.

[Adeyemi Victor Olamide @SnapchatHackwrd]

Cutest tech girl I know @alexabelonix

[Adeyemi Victor Olamide @SnapchatHackwrd]

Anyone who can intro me to an Angel or vc funding??
Would be appreciated

[Adeyemi Victor Olamide @SnapchatHackwrd]

@sridharfyi @16vchq I'm interested