Stay current.
A short weekly digest of the few things that actually matter, connected back to the deeper library.
Starter issue — two essentials
Two foundational reads to anchor the essentials while the weekly digest spins up. From here on, this space surfaces the two things each week that actually change how you should work — nothing else.
- 01Building effective agents
Anthropic
The clearest articulation of 'use the simplest thing that works.' Anthropic separates workflows (predefined, code-orchestrated paths) from agents (the model drives its own steps) and argues most teams should reach for the former. Exactly the language you need when teams are being pushed to 'add agents' and you have to say no well.
- Default to workflows; reserve true agents for genuinely open-ended, hard-to-predict tasks.
- Start from the raw model API and simple, composable patterns — skip the heavy frameworks.
- Every tool is an attack surface: scope it tightly, and design guardrails + observability in from the start.
Related: 🤖 Agents, tools & MCP
- 02Your AI Product Needs Evals
Hamel Husain
The canonical case for eval-driven development. Husain's thesis: your speed of iteration — and therefore your product — is bottlenecked by how good your evals are. The Rechat case study walks the loop of error analysis → tests → judged evals in practice. If you internalize one habit as a lead, make it this one.
- Iteration speed is the whole game, and rigorous evals are what unlock it.
- Layer your evaluation: assertion-style unit tests, model/human-graded evals, and A/B in production.
- Ground everything in error analysis of real traces — not generic public benchmarks.
Curated weekly · paths stay the stable base, these are the live updates