Resources — high-trust sources for Advanced AI Systems

Ground every non-obvious claim in these; cite inline (<sup> → #sources). Populated before teaching. Prefer primary sources (papers, vendor engineering blogs) over secondary commentary. This phase is skeptical by design: most of these sources also tell you when not to reach for the fancy thing.

Primary specs / docs

• Anthropic — “Building effective agents” (anthropic.com/engineering). The anti-hype spine of this track: start with the simplest thing that works; reach for multi-agent only when the task is genuinely parallel and the coordination cost pays for itself. Read its multi-agent caution first.
• Anthropic — “Effective context engineering for AI agents” (anthropic.com/engineering). Long-horizon agents live or die on context budget; this is the reference for compaction and sub-agent context isolation.
• OWASP — Top 10 for LLM Applications (owasp.org). Autonomy raises the blast radius: excessive agency, insecure tool use, and supply-chain risk are the failure modes for the autonomous-systems lesson.

Papers & deeper reading

• Wei et al., 2022 — “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (arXiv:2201.11903). Why a model reasons better when it thinks in steps — the seed idea behind reasoning models and test-time compute.
• Yao et al., 2023 — “Tree of Thoughts: Deliberate Problem Solving with Large Language Models” (arXiv:2305.10601). Search over reasoning paths (explore/evaluate/backtrack) instead of one greedy chain — the bridge from prompting tricks to spending compute at inference time.
• Wu et al., 2023 — “AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation” (arXiv:2308.08155). A concrete multi-agent framework (conversable agents, conversation programming) — the canonical reference for orchestrator-worker and group-chat patterns.
• Park et al., 2023 — “Generative Agents: Interactive Simulacra of Human Behavior” (arXiv:2304.03442). Long-horizon agents with a memory stream + reflection + planning; the cleanest worked example of how persistent state drives behavior over many steps.
• Hu et al., 2021 — “LoRA: Low-Rank Adaptation of Large Language Models” (arXiv:2106.09685). Parameter-efficient fine-tuning: adapt a frozen base model by training small low-rank matrices. The practical default when fine-tuning is actually warranted.
• Hinton, Vinyals & Dean, 2015 — “Distilling the Knowledge in a Neural Network” (arXiv:1503.02531). The original knowledge-distillation paper: train a small student to mimic a large teacher. Background for “distill a cheaper model” vs “fine-tune” vs “prompt.”

Notes on trust

• Vendor blogs describe their product’s strategy; treat product specifics (exact orchestration shape, model names) as version-dependent and verify against current docs before teaching as fact.
• Numbers (token costs, context sizes, latency) drift fast — teach the shape of the trade-off and give concrete numbers only as illustrative “at time of writing.”
• The multi-agent / autonomy space is young and fashion-driven. Where a claim is a vendor’s design choice rather than a settled result, label it as such; prefer the failure-mode framing.