diff --git a/linkedin-posts/personal/agents-picked-wrong-thing.md b/linkedin-posts/personal/agents-picked-wrong-thing.md
new file mode 100644
index 000000000..915c57405
--- /dev/null
+++ b/linkedin-posts/personal/agents-picked-wrong-thing.md
@@ -0,0 +1,41 @@
+---
+title: "I let my agents pick our next product. They were wrong about the wrong thing."
+lane: personal
+pillar: build-in-public
+author: paul-keen
+voice: personal-first-person
+icp_test: Does "the research told us what to build, not whether we'd want to" land, and do founders share their own version of this mistake?
+first_comment: |
+  (value post - no link; reply-CTA only)
+utm_campaign: ""
+utm_content: agents_picked_wrong_thing
+status: draft
+stage: next
+proposed_for: "2026-09-01 17:00 CEST"
+notes: |
+  Pillar: build-in-public (the 70% value; NO course link). From Paul's real work
+  2026-08-24: he ran a multi-agent research sprint (5 parallel agents) to find the
+  best business to build in 2026 with a dev team but no spare millions. Reports came
+  back solid - market size, sources, risk scores. The turn: not one answered the real
+  question (would we want to do it for five years). MCP detail added as concrete hook:
+  we can wire a client's DB/Stripe/Salesforce into an off-the-shelf agent in days, no
+  ML research; MCP ecosystem grew ~900% in 2025 (colrows / CData). Point: do what we
+  already know how to do, package it differently - not chase "idea of the year." Ends on
+  a question, no bait CTA. Plain English, one committed idea, no banned words.
+---
+
+I had my agents run a research sprint this weekend. The question: what business actually makes sense to build in 2026 if you have a dev team but no spare millions to burn.
+
+A couple of hours later, five reports landed. Market size, sources, risk scores - all very solid.
+
+Then I read them and caught the mistake. Not one of those reports answered the real question. They told me where the demand is. None of them told me whether I'd want to do that thing for five years straight.
+
+The tooling makes this fast now. Take MCP - the protocol that lets an AI agent talk to your own systems. We can wire a client's database, Stripe, Salesforce into an off-the-shelf agent in a few days. No machine-learning research, just engineering we already do every week. The ecosystem grew something like 900% last year. Everyone wants agents that touch their own data.
+
+But speed like that doesn't help with the hard part. The hard part is being honest with yourself.
+
+For JetThoughts the answer was both obvious and uncomfortable. Do what we already know how to do - just sell it differently. Not chase the idea of the year. Build the boring connector, ship it, keep the client on a retainer.
+
+Maybe the most underrated skill right now isn't picking up a new stack. It's giving yourself permission to say no to the extra thing.
+
+What are you cutting this year instead of taking on?