Kevin Duigou

I share advice on making good use of AI Agents

#DDD #HexagonalArchitecture #Python #Cpp

[Kevin Duigou @AntifragileFR]

@theiaincameron Incredible it works for India as well
https://t.co/8lI66idTN3

[Kevin Duigou @AntifragileFR]

@theiaincameron UK
https://t.co/8lI66idTN3

[Kevin Duigou @AntifragileFR]

@theiaincameron Incredible Same for France
https://t.co/8lI66ierCB

[Kevin Duigou @AntifragileFR]

@theiaincameron Incredible Same for Spain
https://t.co/8lI66idTN3

[Kevin Duigou @AntifragileFR]

@SergeZaka Il va falloir injecter du Dioxyde de Soufre dans la stratosphere à ce rythme.

Le rapport bénéfice/risque commence à être sans appel!

[Kevin Duigou @AntifragileFR]

@Albirroja @FIFAWorldCup Disgusting Team

Disgusting players

Shame

[Kevin Duigou @AntifragileFR]

@nicolasberrod Le flux d'énergie reçu en Juin dans la cocotte anticyclonique est plus important qu'il ne le serait en Août

La température redescends moins la nuit

le point positif en Juin, c'est pour les gens qui vivent prés de la mer qui n'a pas eu le temps de chauffer

[Kevin Duigou @AntifragileFR]

Most people use Claude Code one prompt at a time.

/goal changes that.

Set a condition like "all auth tests pass and lint is clean". A lightweight model checks after every turn whether it's true. If not, Claude keeps going automatically.

Add auto mode on top: zero interruptions, zero manual re-triggers.

You come back when the work is done.

[Kevin Duigou @AntifragileFR]

/goal in Claude Code is underrated.

You set a condition. Claude loops until it's met.

After each turn, a small fast model (Haiku by default) checks whether the condition holds. Not Claude itself — a separate evaluator. If the condition isn't met, Claude gets the reason and starts another turn.

Pair it with auto mode and you get full autonomy: no tool approval prompts, no manual re-triggers.

You define the finish line. Claude runs until it crosses it.

[Kevin Duigou @AntifragileFR]

Claude Code can now write its own orchestrator on the fly.
Dynamic workflows spawn specialized sub-agents, each with its own context window, model, and worktree.
Here's what this actually solves:

Running a single LLM on a long task produces 3 well-documented failure modes:
Agentic laziness: stops at 20/50 items and declares victory
Self-preferential bias: favors its own outputs when asked to evaluate them
Goal drift: loses initial constraints after context compaction

Workflows fix this structurally.
Each agent has an isolated goal, a clean context, and no visibility into what others produced.
Self-preferential bias disappears when the verifier agent isn't the one who did the work.

The useful patterns:
Fan-out + synthesize: parallelize independent subtasks
Adversarial verification: one checker agent per output
Tournament: N agents on the same task, pairwise judgment until a winner emerges
Loop until done: iterate to a stop condition, not a fixed number of passes

What's most interesting: this goes way beyond code.
"Why did sales drop in March?" → agents on logs, CRM, support tickets, each generates a hypothesis, a panel verifies them.
Multi-agent architecture as an adversarial reasoning framework.

[Kevin Duigou @AntifragileFR]

We spend our lives trying to "find the truth."

ML has a more humble approach:
Define a loss function.
Compute the gradient.
Take a small step in the right direction. Repeat.

It doesn't claim to reach truth. It just gets less wrong over time.

Maybe that's all any of us can do.

[Kevin Duigou @AntifragileFR]

ML taught me something about how to think.

Don't chase the truth directly. You'll get lost.
Instead, define your error. Measure it. Reduce it, step by step.

Gradient descent doesn't know where truth is. It just knows which direction is less wrong.

That's a valid epistemology.

[Kevin Duigou @AntifragileFR]

There is no "truth" in machine learning.

Only loss functions.

Gradient descent minimizes a score. XGBoost stacks residuals.
Neural networks learn by being punished.

[Kevin Duigou @AntifragileFR]

@JLMelenchon L'IA transforme des électrons en « tokens intelligents » créateurs de valeur. Pour rester souveraines et compétitives, la France et l'Europe doivent maîtriser toute la chaîne de valeur : du GW nucléaire au token. Refuser cette réalité, c'est accepter la dépendance.

[Kevin Duigou @AntifragileFR]

@JLMelenchon La souveraineté numérique ne se décrète pas, elle se construit. La France dispose d'un atout unique : une électricité nucléaire abondante et décarbonée. Utilisons-la pour attirer data centers et IA. En s'opposant au nucléaire, LFI affaiblit précisément cette souveraineté.

[Kevin Duigou @AntifragileFR]

Your shell was always composable. Now it reasons.

pytest 2>&1 | tee test.log | claude -p "Analyze failures, return JSON" | jq

cat deploy.log | tee log.txt | claude -p "Summarize errors" > incident.md

[Kevin Duigou @AntifragileFR]

One git alias changed how I review my own code:

git config --global alias.ai-review \
'!git diff HEAD~1 | claude -p \
"Review this diff, output markdown
with Risk Level and Issues." > review.md'

Now git ai-review generates a full report before I open a PR.
I catch my own mistakes before anyone else sees them.

[Kevin Duigou @AntifragileFR]

Claude Code and Codex can handle a lot of complex tasks.

They cannot handle the responsibility of pushing code you actually stand behind.

Understanding what you ship is still yours. That part doesn't get delegated.

[Kevin Duigou @AntifragileFR]

A script without a skill is a tool no one picks up.

The script handles the logic.
The skill handles the judgment — when to trigger it, how to surface the output.

Stop writing one-off automations.
Start building skills that know how to use them

[Kevin Duigou @AntifragileFR]

What's the fastest way to make your AI agents smarter over time without paying more?

Stop letting the LLM rewrite the same logic on every run.

Encode the deterministic parts in scripts.

Let the LLM handle what only an LLM can.

Your skill improves with each iteration. Your token bill doesn't.