https://pygocen.com/blogNotes from building products, fixing what is stuck, and putting AI agents to work.en-gbWed, 23 Sep 2026 09:00:00 GMThttps://pygocen.com/blog/the-model-never-does-the-arithmetichttps://pygocen.com/blog/the-model-never-does-the-arithmeticWed, 23 Sep 2026 09:00:00 GMThello@pygocen.com (Edoardo Pirani)How we let an agent write valuation assumptions and an investment thesis for 32 companies, twice a day, unattended, and why not one number in the workbook comes from the model. Fixed inputs, a schema on the way out, a second computation, and a job that checks the work actually landed.ai agentsguardrailsdatafinancehttps://pygocen.com/blog/silence-is-the-failure-mode-of-alertinghttps://pygocen.com/blog/silence-is-the-failure-mode-of-alertingWed, 26 Aug 2026 09:00:00 GMThello@pygocen.com (Edoardo Pirani)A watchdog died 38 times in a week and every dashboard stayed green. Alerting can only report what someone thought to alarm on, so its failure mode is silence. The five rules we run our systems by, each one paid for by an incident.monitoringalertingreliabilityautomationhttps://pygocen.com/blog/five-layers-of-ai-agent-engineeringhttps://pygocen.com/blog/five-layers-of-ai-agent-engineeringTue, 28 Jul 2026 09:00:00 GMThello@pygocen.com (Edoardo Pirani)Prompt, context, environment, loop, graph. What each layer of AI agent engineering decides, the failure it fixes, when to climb, when to stop, and the two layers the ladder leaves out, from a team that runs a dozen agents on its own products.ai agentsengineeringautomation