Everything about AI agents — how they work, deploy & scale
A teaching resource assembled from ~600 source-backed findings: the mechanisms (how an LLM predicts a token) all the way up to production infrastructure (how you serve, scale, secure and pay for a fleet of agents) — and how every piece combines.
How LLMs work
Tokens, transformers, attention, next-token prediction, sampling, hallucination, context windows.
02How agents work
The agent loop, tool/function calling mechanics, structured outputs, workflow vs agent.
03Reasoning patterns
CoT, ReAct, reflection, tree-of-thoughts, plan-and-execute — with benchmark numbers.
04Context & multi-agent
Context engineering, the four strategies, failure modes, orchestrator-worker systems.
05Knowledge & RAG
Embeddings, vector DBs (HNSW/IVF), chunking, hybrid search, rerankers, GraphRAG.
06Caching everywhere
KV cache, prefix cache, prompt caching (with prices), semantic caching.
07Inference internals
Prefill vs decode, PagedAttention, batching, quantization, GPUs, KV tiering, cost.
08Deploy & scale
KEDA autoscaling, cold starts, durable execution, reliability patterns.
★AI Infra & MLOps
For DevOps → MLOps/AIOps: reference architectures, GPU-on-K8s, company build-outs, roadmap.
09Isolation & safety
Firecracker, gVisor, the isolation ladder, defense-in-depth.
10Memory & state
Short/long-term, episodic/semantic/procedural, MemGPT, Zep, Mem0.
11Real-world use
Coding & browser agents, and case studies (Klarna, Morgan Stanley, LinkedIn).
12Deployments · examples
100+ real deployments across 10 use-cases, each mapped to its infra pattern.
13Ops: eval & security
Evaluation, observability, prompt injection, guardrails, cost optimization.
The landscape
Verified market data: adoption, frameworks, protocols, platforms — the fact-checked report.
INFInfrastructure map
The production reference architecture, managed platforms, and the self-hosted stack.
✗Myths & gaps
Six claims that failed fact-checking, and what the research couldn't confirm.
Confidence: the Landscape pages carry ✅ Verified / 🟡 Reported / ❌ Debunked tags from adversarial fact-checking. The Learn chapters explain established mechanisms, grounded in primary sources (Anthropic, OpenAI, Google research, vLLM, arXiv papers on CoT/ReAct/ToT/Reflexion/MemGPT/PagedAttention) gathered across the research passes.