Field notes from building theagency47, frameworks we use in client engagements, and arguments we want on the public record. New post every 2–4 weeks. Follow via RSS.
theagency47's blog is the long-form publication of founder Christos Papadimitriou, covering AI agent design frameworks, engagement methodology, agency-operations lessons, and arguments about the AI services category in 2026. Posts are published every 2 to 4 weeks. Topics span how-to (specs, evals, monitoring), strategy (pricing, positioning, vertical packs), and field notes from real client engagements where the relevant party has consented to share details. All posts are also distributed via the RSS feed.
theagency47 · Updated May 2026Knowledge workers lose ~1.8 hours a day hunting for information. The 'second brain' idea promises a fix. Here's what the research says about scaling it from one person to a whole company, and where Obsidian fits.
An eval suite is what separates a demo from a deployment. Twenty test cases, four buckets, a clear rubric, automated re-runs on every change. Here's how we structure it.
Five agents, three firms, real production. The pack works, but not in the order you would expect, and three of the agents needed a redesign by month two.
An AI consultant gives you a strategy deck. AI agents give you working software. Both can be the right buy, but only if you know which problem you actually have.
Four frameworks solve four different problems. The right question is not which is 'best', it is which one fits the agent you actually need, with the lock-in profile you can live with.
Returns cost ecommerce brands $10–$20 in labor per ticket and 4–8 minutes of a support rep's day. An AI returns agent handles 75–85% of them end-to-end. Here's the actual workflow.
ChatGPT Team and Claude Projects sit on a salary. A custom AI agent replaces a workflow. The cost curves diverge after month four, and the ROI math is not what most buyers think it is.
Most AI agent failures come from putting the wrong agent on the wrong work. The 3-tier model (Executive, Operational, Task) splits agents by what they're actually good at.