"We trace everything" is an intention, not a guarantee: one new code path is enough to open a hole nobody sees. What holds the rule is a test that freezes the inventory of call sites — plus two traps that cost me time.
"The agent hallucinates" is a bad diagnosis: it sends you looking at the model. On an agent extracting data from documents, I watched the guarantee move — from model quality to three engineering constraints. One of them surprised me.
I kept editing the system prompt, kept re-running, and nothing moved. The culprit wasn't a cache: the graph is built once at startup. Three lifetimes coexist inside an agent, and knowing which one applies changes the feedback loop.
Ask an image model for an App Store screenshot and you get a beautiful visual with an unreadable headline. I inverted the order: a script draws the text, and the model isn't allowed near it. Plus the bit of arithmetic that cost me a full set of screenshots.
An entire access-control design rested on a line of code that doesn't exist. The code wasn't wrong — the ground was. Here's why a fabricated line number is an agent's most effective lie, and the guardrail that catches it.
42 sessions, 913 messages, 155 commits. The report doesn't show what I thought I was doing — it shows what I actually do: hand over broad missions, let them run, and correct mid-flight instead of specifying upfront. Here's the loop, and what it costs.
Perplexity is pulling back from MCP internally, and their reasons are real. But token bloat and weak permissions are a wiring problem, not a protocol one. Here's a POC that exposes my tools once and grants each agent only its slice.
Ti Boug, my iPhone app against the cost of living in Réunion: the real daily prices of fruit & veg, their season, and an AI-composed basket — from the official DAAF price bulletins. What it does, the stack, and where the build stands.
LangChain ships deepagents: the four building blocks I've been repeating for three months, named and packaged. I test them on a real personal need — a banking companion, my statement PDFs kept local, a real chat UI — and find out the pillar that matters isn't the one I expected.
Going from one agent to two isn't about writing twice the code. It's about adding a third piece that never answers: the orchestrator that detects intent and routes to the right specialist.
I'm building an app for Réunion: the real price of fruit and veg already exists, public, stuck in PDFs nobody reads. Ti Boug sets it free — what's cheap, where, and when. The build, and why I challenge the idea before believing it.
Three frontier models, 72 group-stage matches, predictions frozen before kickoff, and a leaderboard that updates itself. A benchmark dressed up as a game — and the perfect excuse to ship a quick mini-app on GitHub Pages.
OpenRouter just shipped Fusion: instead of querying one model, you query several in parallel and a judge synthesizes. The result: three cheap models nearly reach Fable 5's level at half the price. For an indie, that's value right now.
The Astro team ships "The Agent Harness Framework." Flue takes the harness from the previous article and turns it into a real framework. I ported our profile-analysis agent to it to see.
An LLM predicts text; an agent runs tasks. In between sits the harness. I set out to understand it, and realized my mental pattern already described the Claude Agent SDK.
The dev job is shifting from typing to steering: orchestrating agents, writing the spec, guaranteeing that what ships is what was expected. What changes when you stop coding to produce, and start coding to validate.
An AI agent ships «average» code by default: 300-line files, any everywhere, logic in the JSX. Here are the 10 rules I wrote for VoiceJournal, and why they hold.
How I use Claude Skills to brainstorm 10 SaaS ideas scored out of 100, in 15 minutes — with a real run on my local niche, Réunion Island.
Going from Claude that codes to Claude that disciplines itself. SessionStart to bootstrap the repo, PreToolUse to block sketchy commands, Stop to notify you when a task ends. Four concrete hooks running silently in my repos.
A GitHub issue from the phone, a Claude Code session auto-triggered in a cloud container, a PR popping up in Slack, a merge from the GitHub app. My indie workflow when I'm away from the laptop — and what works vs. what doesn't.
Astro 6, Tailwind 4 with @theme, content collections + MDX, bilingual FR/EN, and a full GEO setup (llms.txt, JSON-LD entity graph, robots.txt with allow for AI bots). How and why I rebuilt it all.
Brainstorming → spec → detailed plan → sub-agent execution with review. The workflow I've been using for a few months that finally makes AI usable on multi-day projects, not just snippets.
A Rust CLI proxy that sits between Claude Code and your shell commands. It compresses verbose output (git status, ls, jest...) before it ever reaches the context window. One command to install, instant payoff.
You talk for 30 seconds. Whisper transcribes, Claude writes your journal, gives you a Mind Score and tags your emotions. Here's how it's built and why I did it.
A markdown file at the root of your repo describing how we code here. Read by Claude (the agent) before every session, and by humans at first glance. It's become my most important engineering contract.