I went into the AI engineering courses hoping to feel stupid. I wanted the floor to move under me, the pleasant shock of finding a problem my old instincts could not reach. The first path was called Scope AI Solutions. I was still trying to tell which parts were new when the agentic lab asked for an API key. The project had no credits. I added some. The specified model promptly hit its 50-request daily cap, so the remaining work had to use another model. I had expected to spend the afternoon inside an unfamiliar system. Instead I was troubleshooting a vendor account, which is how I knew I was back in software engineering.

The lesson had a much better name for the work: agentic tool orchestration. A model chooses a tool, supplies arguments, reads the result, and perhaps chooses another. Letting a model make that choice really is new. We have outsourced the if. I still have to decide what it is allowed to call, what arguments it can pass, how long I will wait, and what to record when it stops halfway through. The model gets to be an interesting research problem. The else is me at 2 a.m. in underpants, trying to explain why it refunded a sofa.

Give the workflow another model and it becomes a multi-agent system. Now it has coworkers, roles, and handoffs. Two components can disagree about who sent an email, exactly as they could before. The improvement is that each can write a persuasive account of why the other one sent it. If they cite each other, you have the beginnings of a very confident incident report and still no email. They can, however, schedule a follow-up meeting to align on what “sent” means.

Switching models because of the cap did bring up a question I wanted to spend more time on. The replacement might choose a different tool or give a different answer to the same prompt. The rest of the system would still get a valid response. No compiler would complain. I had changed a dependency in the middle of evaluating the system, and the dependency was the part making decisions. In a human team, replacing the decision-maker would at least call for a handoff; here it was a parameter. One model might ask before issuing the refund. The other might issue two and write a beautiful apology. What would count as equivalent behavior? Which checks would catch the difference? I wanted to sit with that longer than I wanted another name for the routing around it.

Then the retrieval exercise brought me back to permissions. Retrieval-augmented generation gives the model a research assistant: search. It used to hand me three links and let me make a bad decision myself. Imagine it finds an old passage, a correct passage, and someone else’s tax return. The model writes one lovely paragraph out of all three. The index calls them semantically relevant and is absolutely right. The privacy team has a different rubric. Relevance is not permission, freshness, or truth. Search has a mouth now, so in our fantastic present a permissions bug can stand up and tell the whole room what it found.

The RAG lab came back with eight of eleven checks passing and no evaluation errors. Eight passed, three didn’t, and the evaluator itself had worked. I appreciated the distinction. A model answer can be wrong while sounding like it chaired the meeting, and it takes examples and criteria to catch it. The tests were doing real work. I just wished we could admit they were tests. When my old tests failed, they said FAIL and left me to argue with them. Now the model argues back and charges by the token.

I kept looking for the advanced path called Agentic Observability: Who Owns the Mess After the Demo? I would have taken that one twice. Its final lab would give me a refund nobody authorized, a trace with the important bit missing, and three departments pointing at green dashboards. Completion would unlock a lanyard. The lanyard would have access to the logs; I would not.

I went back to the job ads with five badges. I had new failure modes to worry about, but no new engineering instinct. One wanted experience with LLM integration and agent orchestration. An LLM could probably read my résumé and see which years were relevant. The application form had no box for “the same mess, before it could talk.” I think I sent the wrong one of us to school.