Cien Solon 🍋

EU AI Act high-risk deadline got pushed. The technical standards aren't done, so enforcement for high-risk systems is now looking like late 2027, possibly 2028. If you're in a regulated space this is your window. Get your AI governance sorted now .

Today I learned about AI-washing. What is it? Includes: "Undisclosed manual processes masked as AI; AI-themed fraud schemes targeting retail investors; AI systems that, if misrepresented, create cybersecurity or privacy liabilities." Read more here: mcmillan.ca/insights/publi…

@DrorIvry NIST setting the standard early could save everyone from messy retrofits later.

@cienandAI Agent identity will be the sleeper issue. When agents act across services, proving "who authorized this" becomes cryptographic, not just auth. Hoping NIST pushes for verifiable delegation chains - like JWT for multi-hop agent workflows.

NIST has launched the AI Agent Standards initiative last month. It's built on some of these pillars: - industry-led standards - open-source interoperability protocols (so agents aren't locked into one vendor), - and research on agent security and identity

In 2 years an AI agent will have a better transaction history than half the people getting declined for credit cards right now.

The best AI generated comment I've seen on LinkedIn to date.

@thewhychain @thewhychain1 Thanks!

@cienandAI @thewhychain1 dead simple. each agent creates a .lock file when it starts. other agents check for that file before running. if the process dies, the lock has a timestamp and expires after 5 min. no database, no redis, just the filesystem

Today, I built an AI agent that QA’d my entire platform. It signed in as me, walked through every step, found the bugs, and filed feature requests. While it worked, another agent analysed my data and put together a full AEO (AI Engine Optimisation) strategy.

@comforteagle Super useful thanks! And makes sense. DoDs was always my weakness.

@cienandAI Usually breaks when the spec is fuzzy. Before the task runs, write one sentence: done when X is true, failed when Y happens. That's the target the agent checks itself against. Without it, self-verify is just the agent guessing whether you'd approve.

@thewhychain1 How does the file based locks work? Super curious!

@cienandAI running two agents in parallel where one finds bugs while another does analysis, that is the setup. we have 18 running simultaneously across 3 platforms. the coordination problem gets wild once you pass 5, file-based locks ended up being the simplest solution

@comforteagle Need to get better at this. The self verifying process still breaks from my side from time to time

@cienandAI The micromanaging phase usually breaks when you write acceptance criteria before the task, not after. If you can define pass/fail in two sentences, the agent can self-verify and you can skip the review entirely.

So what now? The rest of my thoughts are here: humanandthemachine.substack.com/p/the-day-ai-t…

Until the last six weeks, when the agents got resourceful enough to match it.