Skip to content

Day −1 — Orientation: the 2026 landscape in five reads

Today's focus: orientation before Week 0 · Phase: 0

Must-read (5)

  1. Building effective agents — Anthropic · ~20 min · Agentic AI · The workflow-vs-agent patterns every later topic builds on.
  2. Effective context engineering for AI agents — Anthropic · ~20 min · Agentic AI · Why context engineering replaced "prompt engineering" as the core skill.
  3. What We've Learned From a Year of Building with LLMs — Yan, Bischof, Frye, Husain, Liu, Shankar · ~60 min (skim) · Agentic AI · Tactical, operational and strategic lessons: the production mindset for this roadmap.
  4. The lethal trifecta for AI agents — Simon Willison · ~10 min · Security · Private data + untrusted content + external communication: never give one agent all three.
  5. Designing Data-Intensive Applications, 2nd edition — Martin Kleppmann · ~5 min · System design · The 2e is out in 2026; it's the depth book for Phase 5.

Worth knowing

  • Spring AI 2.0 went GA on 12 Jun 2026 (Boot 4, @McpTool annotations, Streamable HTTP for MCP).
  • The MCP 2026-07-28 spec revision (RC) moves to a stateless core and adds Tasks and Extensions.
  • Microsoft Agent Framework 1.0 (Apr 2026) supersedes AutoGen and Semantic Kernel for new work.
  • OTel GenAI semantic conventions are still at Development status: adopt them, but pin versions.

Question of the day

L3 — Your team wants a multi-agent system for customer-support triage. What do you ask before agreeing?
Answer

Start from the task, not the architecture:

  • Can a single LLM call plus retrieval, or a fixed workflow (prompt chaining or routing), solve it? Anthropic's guidance is to add agentic autonomy only when a predefined path can't handle the variety of inputs.
  • Ask for the eval set first: 50–100 real tickets with expected outcomes.
  • Ask about the cost and latency budget per ticket.
  • Ask which tools need write access. Irreversible actions need human approval, and you need to check for the lethal trifecta.
  • Ask how failures are observed (tracing).

Multi-agent adds coordination overhead and new failure modes, such as agents looping or losing context at hand-offs. Justify it with measured gains over a single agent on the eval set.

Tip for this weekend

Do the Week 0 baseline before reading ahead. The skip list it produces will save you hours in Phase 1.

English word of the day

Trade-off (noun; also to trade off) · Register: neutral, standard in engineering. · Meaning: a situation where you accept less of one good thing to get more of another. · Example: "The main trade-off is consistency versus latency." · Style tip: the noun is usually hyphenated (trade-off); the phrasal verb is two words (we traded off latency for consistency). Prefer trade-off between A and B to trade-off of A and B.