# Agentic Design

> A multilingual reference catalog for AI agent architecture, implementation patterns, trade-offs, and production practices.

Full catalog with per-technique summaries and canonical URLs: https://agentic-design.ai/llms-full.txt

Content is available in English, French, German, Japanese, Simplified Chinese, and Russian. Technical product and pattern names are retained where translating them would reduce precision.

## English
- [Home](https://agentic-design.ai/)
- [AI agent pattern catalog](https://agentic-design.ai/patterns)
- [Model architectures](https://agentic-design.ai/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/fine-tuning)

## Français
- [Home](https://agentic-design.ai/fr)
- [AI agent pattern catalog](https://agentic-design.ai/fr/patterns)
- [Model architectures](https://agentic-design.ai/fr/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/fr/fine-tuning)

## Deutsch
- [Home](https://agentic-design.ai/de)
- [AI agent pattern catalog](https://agentic-design.ai/de/patterns)
- [Model architectures](https://agentic-design.ai/de/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/de/fine-tuning)

## 日本語
- [Home](https://agentic-design.ai/ja)
- [AI agent pattern catalog](https://agentic-design.ai/ja/patterns)
- [Model architectures](https://agentic-design.ai/ja/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/ja/fine-tuning)

## 中文
- [Home](https://agentic-design.ai/zh)
- [AI agent pattern catalog](https://agentic-design.ai/zh/patterns)
- [Model architectures](https://agentic-design.ai/zh/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/zh/fine-tuning)

## Русский
- [Home](https://agentic-design.ai/ru)
- [AI agent pattern catalog](https://agentic-design.ai/ru/patterns)
- [Model architectures](https://agentic-design.ai/ru/model-architectures)
- [Fine-tuning open-weights models](https://agentic-design.ai/ru/fine-tuning)

## Field notes (English only)
Mechanism teardowns: a framework abstraction opened up and rebuilt from scratch with no dependencies.
- [Why your agent bill grows quadratically](https://agentic-design.ai/field-notes/why-agent-bills-grow-quadratically): An agent resends its whole transcript on every step, so a run that takes twice as many steps costs about four times as much. The arithmetic, and the three fixes.
- [What an Agents SDK handoff actually does](https://agentic-design.ai/field-notes/what-a-handoff-actually-does): A handoff looks like transferring control to another agent. Underneath it is a tool call whose result is a different system prompt and a different tool list.
- [Structured outputs, and the bug they cannot fix](https://agentic-design.ai/field-notes/structured-outputs-and-what-they-cannot-fix): Constrained decoding guarantees your JSON parses. It guarantees nothing about whether the values are right, and the failure it hides is worse than the one it removes.
- [The retry that charged twice](https://agentic-design.ai/field-notes/the-retry-that-charged-twice): Agent frameworks retry failed tool calls by default. If the tool moved money, sent mail or created a record, the retry did it again. Idempotency keys, from scratch.
- [What create_react_agent actually does](https://agentic-design.ai/field-notes/the-agent-loop-under-every-framework): Every agent framework ships a prebuilt ReAct agent. Underneath all of them is one while loop, about forty lines of TypeScript, no dependencies.
- [LangGraph's checkpointer from scratch: what state persistence really costs you](https://agentic-design.ai/field-notes/checkpointers-and-what-state-costs): A checkpointer is a dictionary with a write on every step. Building one takes twenty lines; the bill arrives as write amplification and a serialization format you now have to version.
- [CrewAI's hierarchical process is a supervisor pattern. Here it is in one file.](https://agentic-design.ai/field-notes/hierarchical-process-is-a-supervisor): Swapping one enum value changes who decides what runs next. The mechanism behind it is a manager model choosing among workers, and it predates every framework that ships it.
- [MCP tool calling without the SDK](https://agentic-design.ai/field-notes/mcp-tool-calling-without-the-sdk): The Model Context Protocol is JSON-RPC over a pipe, and since the 2026-07-28 revision there is no handshake at all. A whole server fits on one screen, and seeing the wire makes the security surface obvious.

## Model Context Protocol (English only)
Reference for MCP at protocol revision 2026-07-28. Each page states the revision it describes: this protocol has removed a handshake, sessions and a streaming endpoint, so undated material about it is unreliable.
- [Model Context Protocol](https://agentic-design.ai/mcp): hub.
- [MCP Transports: stdio and Streamable HTTP](https://agentic-design.ai/mcp/transports): The two transports MCP defines, how each frames messages, and what the 2026-07-28 revision removed from the HTTP one: sessions, the GET stream, and resumability.
- [Building an MCP Server](https://agentic-design.ai/mcp/servers): What a server owes a client beyond tools: discovery, resources and prompts, cacheable lists, error conventions, and shutdown that does not need a kill signal.
- [Consuming an MCP Server You Did Not Write](https://agentic-design.ai/mcp/clients): Version negotiation without a handshake, detecting whether a server is modern or legacy, timeouts, cancellation, and the trust boundary you cross by connecting.
- [MCP Security: the Threat Model](https://agentic-design.ai/mcp/security): Every model-facing string in MCP is an injection surface, a stdio server is a local process with your privileges, and the protocol enforces none of it. What that means in practice.
- [MCP Versioning: Modern and Legacy Eras](https://agentic-design.ai/mcp/versioning): The 2026-07-28 revision split MCP into two eras. What moved, which combinations of client and server actually work, and how to keep a page about this protocol from going stale.

## Agent benchmarks (English only)
Twelve agent benchmarks ordered by the question each one answers, each with what a good score on it still does not prove. No scores and no ranking: these benchmarks measure different things, and leaderboard numbers date faster than the page could track them.
- [Agent benchmarks compared](https://agentic-design.ai/benchmarks): 15 benchmarks across 7 questions, each linking to its paper and its pattern page.

## Who we follow (English only)
- [Who we follow in AI](https://agentic-design.ai/sources): every publisher behind the news hub, grouped into the labs and model providers, the serving and tooling teams, the researchers writing under their own name, and the aggregators. Each one links to its own site and to its stories here.

## Patterns that look alike (English only)
Pairs that get chosen wrong because their names are close: the one axis that separates each pair, the condition for each side, and what the wrong choice costs.
- [Both of these break a large task into smaller ones. Which one do I want?](https://agentic-design.ai/compare#htn-vs-goal-decomposition): Whether the breakdown already exists. HTN executes a decomposition somebody encoded in advance as domain methods. Goal decomposition produces a decomposition for an objective that arrived vague.
- [Both route work to a more specialised agent. What is the difference?](https://agentic-design.ai/compare#supervisor-vs-handoff): Who holds control afterwards. A supervisor delegates and keeps it, fanning work out and collecting results back. A handoff gives control away and does not expect it back.
- [These two names look interchangeable. Are they the same thing?](https://agentic-design.ai/compare#chat-vs-conversational): Who opens the exchange, and whether the medium is fixed. Chat is a message log the user drives, and its problems are threading, context and rich content. Conversational covers an agent that may speak first and may not be speaking in text at all.
- [In the loop, on the loop. Is there a real difference or is it jargon?](https://agentic-design.ai/compare#human-in-vs-on-the-loop): Whether the system waits. In-the-loop blocks: nothing proceeds until a person decides. On-the-loop runs at full speed and a person watches, stepping in when something trips an alert.

## Publisher
- Agentic Design is published by KORTEXYA SAS, France.
- Contact: contact@kortexya.com