Go Deeper

Everything here earned its slot; if something was merely good, it’s not listed. Each entry says why it’s worth your time and who it’s for. Where the field disagrees, both sides are listed on purpose. Links go to sources that move — if one has drifted, the title plus author will find it.

The agent loop

• Building Effective Agents (Anthropic) — the essay that fixed the field’s vocabulary: workflows vs. agents, the composition patterns, and the “simplest thing that works” ethos this hub inherits. Read first, before any framework.
• How to Build an Agent (Thorsten Ball) — builds a working coding agent in a few hundred lines, live, no framework. The fastest cure for the belief that agents are magic; ideal right after the Agent Loop Apply-it.

Context engineering

• Effective Context Engineering for AI Agents (Anthropic) — the practitioner playbook this hub’s topic compresses: altitude, just-in-time retrieval, compaction, memory. For when you’re building and need the full detail.
• Context Rot (Chroma) — the measurements behind “capacity isn’t attention”: performance degrading with input length across models, including where vendors’ needle-in-a-haystack marketing says otherwise. For the empirically minded.
• “How Long Contexts Fail” (Drew Breunig, dbreunig.com) — a crisp taxonomy of the failure modes — poisoning, distraction, confusion — with fixes for each. The best short read for naming what you’re seeing in a misbehaving trace.

Tool design

• Writing Effective Tools for Agents (Anthropic) — the craft guide: consolidation, namespacing, token-efficient returns, and evaluating tools with agents in the loop. The full version of this hub’s topic.
• Model Context Protocol (spec and docs) — the standard itself, with server-building tutorials. Go here when your capstone involves shipping a tool anyone else can connect to.

Evals

• Your AI Product Needs Evals (Hamel Husain) — the essay that made rigorous evals a mainstream practice: error analysis, looking at your data, judge calibration. For anyone about to build their first suite.
• What We Learned from a Year of Building with LLMs (Yan, Bischof, Frye, et al.) — hard-won operational lessons from six practitioners, evals and beyond. Best read after you’ve shipped something and recognize the wounds.
• Patterns for Building LLM-based Systems (Eugene Yan) — the wider pattern language around evals, guardrails, and monitoring, with references worth mining. For the systems thinker who wants the map behind the practice.

Multi-agent systems

Read these two together — the disagreement between them is the current state of the art:

• How We Built Our Multi-Agent Research System (Anthropic) — the candid production case for orchestrator–worker: where it won big, and the honest token bill. The best public account of handoff engineering at scale.
• Don’t Build Multi-Agents (Cognition) — the case against: why shared context beats delegation for most tasks and how handoffs lose the plot. If this essay doesn’t sting a little, reread your own architecture.

Cost and latency

• Prompt caching (Claude docs) — the mechanics this hub’s topic describes: breakpoints, TTLs, write vs. read pricing, and the invalidation rules. Keep open while doing the caching Apply-it.
• Prompt caching guide (OpenAI docs) — the other major provider’s automatic-caching model. Worth skimming for the contrast: what’s a breakpoint decision on one platform is a prefix-stability discipline on both.

Staying current

The field moves monthly; curation beats archives.

• Simon Willison’s blog — the field’s most reliable filter: hands-on, skeptical, fast, and honest about what he got wrong. If you follow one source, this one.
• Latent Space (podcast and newsletter) — long-form interviews with the people building the tools and models. For depth on where the frontier thinks it’s going.