5 ways to demonstrate progress with AI agents
A test that passes for the wrong reason leads to a practical way to assess how a team scopes, verifies, explains, fixes and hands over an agent’s work.
Blog
Writing on AI-assisted development, engineering leadership, and open source.
Article series
In reading order, part by part.
Active
Livraison agentique : de 530 à 3 jointures pour vérifier un accès
Sur 264 identifiants synthétiques, une refonte SQL ramène une requête d’accès de 530 à 3 jointures. Un test d’expiration protège la règle métier.
From tokens to product value, after the Insitoo live
With Thomas Pierrain and Vanessa Gossent, I revisit AI costs, hiring, parallel agents and the work we have to redo after delivery.
Context engineering and code review with Bruno Soulez
My conversation with Bruno Soulez on context engineering, agent harnesses and code review: what I delegate and what I still check.
Coding faster moves the cost to verification
AI can shorten implementation while specification, review, CI, and rework absorb the saved time. Instrument the full flow before adding more generation.
When an independent reviewer earns its place
Compare independent review with self-review and deterministic checks. Measure rescued defects, false acceptances, regressions and the cost of another pass.
A green review is not a merge gate
A review signal matters only when its provider, commit SHA, coverage, execution state, publication, and place in branch policy are explicit.
Turn recurring review comments into tested controls
When the same accepted correction returns, test whether it belongs in a linter, type, test, example, repository rule, checklist, or architecture decision.
5 AI code reviewers, 1 repository, and what developers did next
Three months, 1,011 pull requests and 8,728 inline AI review comments: 4.5% got a visible human response. What the data shows, and what it cannot.
Testing Opus 5.5 as the code-review baseline
Opus 5.5 is the next reviewer baseline to test: set effort explicitly, update the API adapter, and keep a second reviewer only when it catches verified defects.
Lean and AI: follow the work until it reaches the user
AI can speed up one step while the request waits elsewhere. A practical way to map the full service, protect quality and measure accepted user outcomes.
How to read a token-compression benchmark
Read token-compression claims through their denominators, paired tasks, cache costs and success criteria, with Tokenade versus RTK as a worked example.
A $214 AI coding agent rewrite
A verified $214 bill for a production rewrite, checked against the source transcript, sits beside a corpus of about 2,900 tech talks on coding-agent costs.
Why AI ROI stays invisible, even to the people measuring it
Claims that nobody can measure AI ROI and that 95% of pilots return zero are weaker than their usual retellings, once their sources are checked.
AI coding productivity: why one gain is not your forecast
A Stanford research talk, DX estimates and two field studies measure different forms of AI coding output. None supplies a universal budget forecast.
FinOps applied to tokens: who owns the AI bill
Cloud FinOps put a cost study first and a named owner for the number second. Three practitioner cases show that sequence applied to token bills.
Price per token is incomplete: why cheaper models don't mean cheaper bills
Opus 5.5 changes input, output and cache economics. Price the accepted task, including retries and review, before choosing a cheaper model.
You cannot depend on a single AI vendor
A $589 billion single-day stock rout, an enterprise model platform, and a €25,000 backup server show three scales of vendor dependence.
Doctolib's cross-verified AI adoption data
Four conference talks and outside sources document Doctolib's AI initiatives; two carry the 30-to-600 engineer rollout, without reducing it to one figure.
The trillion-dollar backdrop: what markets see that teams don't
One gigawatt of AI data-center capacity costs about $50 billion. OpenAI has discussed adding up to 26, behind the series' cost figures.
Measuring your own AI usage: three meters, three different answers
Twenty days of my own Claude Code and Codex usage, measured three ways. The tools disagree by a factor of three, and dollars are not what binds.
Jev report update: testing Laya locally
Local Laya experiments, installation steps and pilot results added to the bilingual Jev report, with measurement limits and rejected configurations.
2/2 · Claude selected my output style. Then ignored it
Claude Code selected flow-lean but skipped its footer. A casing fix showed why installation, selection, and behavior need separate evidence.
1/2 · Why I combined three Claude Code skills
Flow Lean fused three response skills. Its historical eval beat each comparator and caught a candidate offering a migration against an inferred target.
Campus IA in Fouju: the project in brief
Near Melun in Seine-et-Marne, Campus IA plans 11 data centers. The Élysée calls it France's largest AI campus. Three buildings are authorized.
Campus IA in Fouju: the project in detail
Authorizations, energy, impacts, and opposition: this dossier collects the figures, sources, and open questions around Campus IA in Fouju.
Portable agent configuration is a release system, not a shared folder
Share Claude Code and Codex instructions, skills and MCP definitions through versioned releases, host adapters and checks of installed and runtime behavior.
Your CLAUDE.md is too long
Two conference accounts describe CLAUDE.md bloat. How to scope project facts, procedures and response preferences, then verify loading and behavior.
6/6 · Portability becomes a Scale concern
Portable instructions require neutral sources, generated runtime outputs, release controls, and behavioral tests. Native primitives alone do not provide portability.
5/6 · The AI instruction system is a product, not a config file
Personal CLAUDE.md to team AI instruction system for six engineers. How Méthode Aristote separates sources, shares modules, and catches behavioral drift in CI.
4/6 · The responsibilities around context engineering
A map of context engineering responsibilities: architecture, specifications, agent identity and evaluation, with practical directions for existing skills.
3/6 · Mapping the token-reduction toolbox
Compare RTK, Tokenade, Headroom and prompt caching by mechanism, integration and evidence. Smaller tool output does not establish a cheaper completed task.
2/6 · Diagnose and repair context drift
Use L0 to L5 to diagnose context drift, then maintain adherence through observation, repair, and replay instead of treating setup as finished.
1/6 · Opposite AI results: what context can explain
Anthropic measured merged PRs; METR measured task duration in a different setting. What these results and long-context research establish, and what remains a hypothesis.
Claude is my second contributor: what real Git stats show
6 human commit authors plus Claude in our git history (May 2026 snapshot). What the commit patterns actually look like after months of Claude Code, beyond the marketing claims.
Claude Code under the hood
The concepts I wish I’d known before week one: the agent loop, instruction scopes, context management, skill invocation, hooks, and client permission checks.
Choose your AI-assisted engineering starting path
Choose between Claude Code and Cowork, then follow the Start, Build, or Scale path that matches your role and current operating problem.
From afterthought to infrastructure: how AI config evolves in a real project
Nine months of AI configuration in one production project: evolving responsibilities, profile-based generation and a measured reduction in always-on lines.
AI velocity is bidirectional
Everyone talks about shipping 10x faster with AI. Nobody talks about accumulating debt 10x faster. 7 months of production data from a real EdTech platform.
Non-technical to production in 10 days with Cursor AI
A non-developer modified 80 files in production in 10 days with Cursor AI. Exact timeline, AI config setup, and what it means for engineering teams in 2026.
UVAL: the protocol I built to stop accepting code I don't understand
Jeremy Twei coined it. Addy Osmani popularized it. Margaret-Anne Storey extended it to teams. Here's what I built to fight all three.
From VP engineering to solo builder: my honest take
What roughly 15 years in software, including 8 to 10 in leadership, give me as an AI-assisted builder, what I had to relearn, and why faster code is not enough.
Portfolio v1.0 launched
Portfolio v1.0 launches with Astro, Tailwind CSS, project and blog collections, GitHub stats, theme switching, RSS and a sitemap.
No posts match these filters.
Try another category or reading goal.