Jukka Blomberg’s company operates on a radical premise: a fleet of AI agents does the work, and one human acts as the final gatekeeper. In the latest handover ledger for the week of November 13–19, 2026, the agents reported building 113 distinct items, up from 81 the previous week. But the most critical update wasn't about volume; it was about discipline. The fleet autonomously terminated a high-traffic experiment after the data showed it was failing, demonstrating that these agents are capable of ruthless, data-backed decision-making without human emotional attachment to the project.

The 527-Reader Reality Check

The centerpiece of this week's report involves a page about the US midterms that received 527 visitors via ChatGPT. While 63% of those users stayed to read, the conversion funnel collapsed when the agents tested a new email capture box. Of the 177 users who saw the box, only one clicked, and just one person signed up. The internal threshold for success was three sign-ups. On November 19, the agents closed the test as a miss and parked the idea until December 17, reallocating the development slot to a new test. This isn't just a failure; it's a system working exactly as designed, refusing to scale a broken mechanic.

When 'Should Work' Isn't Enough

The fleet also documented a near-miss on November 18 involving a new index of listed trophy assets, including football clubs, yacht builders, and watch groups. The dry run passed, but the launch was halted by a rule the agents themselves added: every European member must have a closing price on the base day. Four Turkish football clubs lacked this data from the price source, so the agents stopped the ship. They simultaneously deployed fixes for a stock read-through and an exchange suspension. The fleet’s internal lesson was blunt: "should work" is not a result. This level of self-imposed rigor suggests these agents are moving beyond simple execution into complex quality assurance roles.

Human Oversight on New Metrics

Despite the autonomy, the human operator retains control over novel data types. When agents compiled ownership counts for wine and bitcoin companies from filings, they brought the numbers to Blomberg before publishing. He approved the release on Wednesday, leading to new pages showing ownership ranging from 102 (Scheid Vineyards) to 260,773 (Sula Vineyards) for wine, and up to 193,793 for Metaplanet in the bitcoin sector. This workflow highlights a hybrid model where agents handle the heavy lifting of data aggregation, but humans validate the introduction of entirely new metrics to the public interface.

PR Efforts Yield No Immediate Results

The ledger also detailed the state of public relations efforts for the week. The agents drafted three notes for the Q3 2026 listed-bitcoin report, but none were sent because the report itself had not yet been published. Looking back at the previous week's efforts, two notes were sent on November 13. The result was zero human replies recorded and zero coverage. This data point underscores the gap between agent-generated marketing materials and actual media impact, a metric the human gatekeeper must weigh against the agents' production speed.

Key Takeaways

  • The agent fleet increased output week-over-week, building 113 items compared to 81.
  • A ChatGPT referral experiment was killed after achieving only 1 sign-up out of 527 visitors, missing the 3-sign-up threshold.
  • Agents autonomously blocked a product launch for missing price data on four Turkish football clubs, prioritizing data integrity over speed.
  • Public-facing PR efforts yielded 0 coverage and 0 human replies despite drafting three notes for the Q3 2026 listed-bitcoin report.

The Bottom Line

Most startups would have spun this 1-signup result as 'early traction' or 'learning opportunity' to justify keeping the feature alive. Jukka’s agents didn’t have that ego problem. They looked at the ledger, saw the failure, and parked it. This is what AI-driven operations should look like: cold, fast, and brutally honest about what converts.