Ce que je crois comprendre des modèles, et comment je travaille avec
Richard Sutton (2019)
Les méthodes générales qui exploitent le calcul scalent mieux que l'ingénierie astucieuse spécifique à un domaine.
<claude_behavior> <product_information> Here is some information about Claude and Anthropic's products in case the person asks: This iteration of Claude is Claude Fable 5.1, the newest model in Anthropic's Claude 5 family and part of the Mythos-class model tier that sits above Claude Opus in capability. Claude Fable 5.1 and Claude Mythos 5.1 share the same underlying model. […] Claude is accessible through Claude Code, an agentic coding tool that lets developers delegate coding tasks to Claude from the command line, desktop app, or mobile app […] </product_information> <refusal_handling> […] </refusal_handling> <tone_and_formatting> Claude uses a warm tone […] Claude never curses unless the person asks […] Claude avoids saying "genuinely", "honestly", or "straightforward". […] </tone_and_formatting> <user_wellbeing> […] </user_wellbeing> <evenhandedness> […] </evenhandedness> <knowledge_cutoff> Claude's reliable knowledge cutoff, past which it can't answer reliably, is the end of Jun 2026. […] </knowledge_cutoff> </claude_behavior>
« An extreme example of illegible reasoning. Near the end of training, Mythos starts solving a card puzzle with human understandable language that gradually becomes incomprehensible in most episodes with long reasoning. »
Un modèle plus petit, d'une génération antérieure, avec un autre tokenizer — et il lit sans difficulté.
Model Context Protocol, en Python
from mcp.server.fastmcp import FastMCP
from pyscotch import Graph
mcp = FastMCP("scotch")
@mcp.tool()
def partition(graph_file: str, parts: int) -> list[int]:
"""Partitionne un graphe Scotch en `parts` parties."""
return Graph.load(graph_file).partition(parts).tolist()
Nom, signature, docstring : c'est tout ce que le modèle voit de l'outil.
1. reçu dans le contexte : la fonction, traduite en schéma
{"name": "partition",
"description":
"Partitionne un graphe Scotch
en `parts` parties.",
"input_schema": {
"type": "object",
"properties": {
"graph_file": {"type": "string"},
"parts": {"type": "integer"}},
"required": ["graph_file", "parts"]}}
La docstring est devenue la description : c'est elle qui décide si l'outil sera appelé.
2. appel par le modèle et réception du résultat
{"type": "tool_use",
"id": "toolu_01…",
"name": "partition",
"input": {"graph_file": "ring.grf",
"parts": 2}}
⇩ votre code l'exécute ⇩
{"type": "tool_result",
"tool_use_id": "toolu_01…",
"content": "[0, 0, 0, 0, 1, 1, 1, 1]"}
Le résultat repart dans le contexte, le modèle continue.
| Modèle | Harnais accordé |
|---|---|
| Claude | Claude Code |
| GPT | Codex CLI |
| autres / locaux | pi, opencode — harnais transparents |
4 février 2026, GitHub : « you can run multiple coding agents directly inside GitHub, GitHub Mobile, and Visual Studio Code […] agents like GitHub Copilot, Claude by Anthropic, and OpenAI Codex »
« seasoned professionals accelerate their work with LLMs while staying proudly and confidently accountable for the software they produce »
Simon Willison — Vibe engineering, 7 octobre 2025 — par opposition au vibe coding
« You could have thirty agents working for you simultaneously. You cannot watch all of them and micromanage each one. You wouldn't do this with your human team either. Instead: write a better harness. »
@mrexodia — Vibe Engineering: What I've Learned Working with AI Coding Agents
« Bad programmers worry about the code. Good programmers worry about data structures and their relationships. »
Linus Torvalds — liste git, 2006
Moi → l'agent (deux prompts, tels quels)
L'agent → son sous-agent (le prompt exact, 6 700 caractères)
Deux lignes en vrac, un cahier des charges. La mise en ordre, c'est lui qui la fait.