Most “agentic AI” is just if-else trees wearing a trench coat. Real agency sets goals, recovers from faceplants, and knows when to call a human.
The story
Patterns:
– “X, not Y” repeats.
– Punchlines feel too tidy.
– Tool list reads roundup-ish.
– Forecast needs more human texture.
– Ending is slogan-y.
Rewrite:
Most “agentic AI” workflows I’m seeing aren’t agentic.
They’re if-else trees in a trench coat made of GPT calls.
Agency means the system sets small goals, recovers after a faceplant, and knows when to tap a human. Not “panic, page Dave, call it learning.”
What I keep seeing:
– Prompt chain A hands off to B
– Output gets weird
– Human duct-tapes it
– Loop respawns tomorrow
That’s anxious automation. Not agency.
It matters because software work is shifting. Copilot, Cursor, and Claude Code move humans from typing code to describing intent. Yay for wrists.
But SaaS has an awkward zit forming. If a custom tool can show up in an afternoon, per-seat generic trackers and form builders start feeling like fax machines with logins. Network effects, owned data, and regulatory armor still count. The soft middle gets thin.
Then agents walk in: Operator, Perplexity Assistant, Claude Computer Use. We move from “tool I operate” to “agent I instruct.” The app helping me do the thing now competes with the agent doing the thing. Expense app, meet: “watch my expenses and flag weird stuff.”
So the question I’m sitting with: is this product a toll booth between intent and outcome? Toll booths get squeezed.
Memory beyond the chat window. Tool use without babysitting. Failure plans made before 2am production gremlins arrive. That’s where the moats live.
I wrote the longer rant, “The Extinction of Software,” on clearsay.net if you wanna chew on it.
What if the move is less anxious automation, more agency with a brain?
Recent Comments