tac

The Agent Stack — a technical reference for Agentic AI.

srmdn.github.io/tac

---

What is this?

tac is a structured reference that maps the full Agentic AI stack — from token economics and quantization up to multi-agent orchestration and production evals.

It is not a tutorial. It is a reference for builders: production decision framing at the top of every page, technical depth throughout, and a "Production Reality" section on each topic with real gotchas.

The Stack

┌─────────────────────────────────────────────────────────────┐
│  Agent Applications   ← coding agents, assistants, support  │
├─────────────────────────────────────────────────────────────┤
│  Agent Runtime        ← tool loops, MCP, memory, planning   │
├─────────────────────────────────────────────────────────────┤
│  Agent Infrastructure ← serving, caching, routing, observ.  │
├─────────────────────────────────────────────────────────────┤
│  Foundation Models    ← transformers, tokenizers, weights   │
└─────────────────────────────────────────────────────────────┘

Topics

Foundations

• Tokens & Cost — pricing tables (6 providers), language ratios, cost math
• Context Windows — model sizes, lost-in-the-middle, RAG vs large-context
• Sampling — temperature/top-p mechanics, structured output APIs, temp=0 non-determinism

Infrastructure

• LLM Serving — vLLM, SGLang, TGI, TensorRT-LLM, managed vs self-hosted
• Prompt Caching — Anthropic/OpenAI/Google/DeepSeek breakdown, break-even math
• Latency — TTFT/TPOT, latency budget table, streaming, parallelization
• KV Cache & Quantization — KV memory math, PagedAttention, GPTQ/AWQ/GGUF
• Rate Limits & Concurrency — tier tables, TPM math, backoff/queue/circuit-breaker

Architecture

• MoE Architecture — router mechanics, self-hosting paradox, load balancing
• Reasoning Models — thinking tokens, cost reality, hybrid routing

Agents

• Agent Frameworks — 2025 landscape (LangGraph, CrewAI, AutoGen), MCP adoption, raw-loop example
• Orchestration — 4 patterns with tradeoffs, failure contracts, consensus anti-pattern
• Tool Use & Function Calling — mechanics, reliable tool design, argument hallucination
• Context Management — sliding window, summarization, RAG, token budgets

Quality & Safety

• Evaluations — eval types, LLM-as-judge gotchas, harness design, eval drift
• Prompt Injection & Security — attack surface, mitigations, system prompt leakage

Built with

• VitePress — static site generator
• TypeScript + CSS — theme and config

Author

Said Ramadhan (@srmdn)

License

MIT