Runtime Control, Agents & Execution

A Hard Truth About “Agentic AI” That Keeps Getting Dodged

The moment you decide you will allow an agent or system to read, modify, create, or execute, you have granted authority and you have liability.

Chris Ciappa· January 19, 2026· 3 min read

A Hard Truth About “Agentic AI” That Keeps Getting Dodged

The moment you decide you will allow an agent or system to read, modify, create, or execute, you have granted authority and you have liability.

Member article

Full article access requires Synth Membership

This article is available to Synth members. Membership is $50 per year and unlocks member articles through your Samirac account.

Become a Member Sign in

Related work

Browse all →

I BUILT A CONTROL AROUND CHATGPT BECAUSE TELLING IT WHAT I WANTED WASN'T ENOUGH

I gave ChatGPT an external specification for how I write. It understood it, had followed it before, and then ignored it anyway. What happened next became an unexpectedly clean demonstration of AI drift, external correction, and why having a rule is not the same thing as enforcing it at execution.

Gaia Nexus and Samirac AI Demonstrate a Working Interoperability Model

Samirac AI and Gaia Nexus completed a live interoperability exercise inside dAIsy using Gaia Nexus Detection Events and a Samirac Execution Contract. The test demonstrated that external evidence can change execution requirements—from admissible, to additional validation, to second review—without changing the actor’s identity, authority or permissions.

Did You Ever Notice the Creep in Medicine?

And by creep, I mean drift or institutional drift.