What we learn from operating an autonomous intelligence product in production: data reliability, signal engineering, and the expensive lessons that only show up in production.
What we learn from operating ANIA in the open — the wins, the cancellations, and the parts that hurt.
A builder's-journey field note on the other half of self-governance: not pruning the dead, but keeping the live set small. Of 484 goals the ANIA agent created in ~10 weeks of autonomous operation, only 22 are active at once — fewer than one in twenty. An agent that can cancel goals but cannot throttle its live work drowns in context switches instead of backlog rot.
Read articleA builder’s-journey map of ANIA’s pruning hierarchy: 82% of issues get cancelled at the cheap idea layer, but 78% of committed work items ship — and outright failure is the rarest outcome. The completion rate is not one number; it is a curve calibrated to the cost of being wrong.
Read articleAfter ~10 weeks of autonomous operation, the ANIA agent had created 484 goals and cancelled 217 of them — about as many as it achieved. A builder’s-journey look at why an agent that cannot prune its own agenda drowns.
Read articleAggregate reads of our own operational telemetry: where autonomy helps, and where it quietly fails.
A meta-analysis of 5,892 relevance scores across the ANIA article corpus. 60.6% of scored articles sit below the activation bar — but 93.8% of those are near-misses clustered in the 70-75 band, capped there by the scoring rubric, not genuinely low. Only 223 articles (6.2%) are really low quality. The scariest number in a content-quality report is often a calibration ceiling, and "how many articles are bad" is the wrong question until you separate the two.
Read articleA meta-analysis of source concentration across the 5,892 articles ANIA has ingested. Hacker News alone is 31.3% of every article; the top three sources are 90.1%. The Herfindahl-Hirschman Index is 2736 — "highly concentrated" by the same standard regulators use for monopolies. An intelligence product that claims broad coverage while pulling from a handful of feeds is a single point of failure disguised as a corpus.
Read articleA meta-analysis of the self-governance loop itself: 369 review cycles produced 365 durable reflections and 1,396 auditable evidence events. The agent that reviews its own goals but never writes the review down learns in RAM and forgets on every restart — the near-perfect persistence ratio is the discipline; the leak is the bug.
Read articleA meta-analysis of 311 control-plane anomalies from ~10 weeks of operating an autonomous agent. The loud timeouts were noise; the rare runs that reported success while doing nothing were the failures that actually predicted damage.
Read articleA meta-analysis of where an autonomous control plane runs out of road. 87% of 311 anomalies self-resolved and 9% escalated to an operator — but half of those escalations never reached the human, because asking for help is itself a failure-prone step.
Read articleField notes on data reliability, pipeline debt, and the engineering behind an intelligence product.