Jonathan Gelin

Jonathan Gelin, AI harness engineer for developer experience

Coding agents are becoming a commodity. Adapting your teams and your SDLC is the hard part.

That is the part I help with: the strategy, and the harness that connects agents to your standards, your CI/CD and your data.

Book a 30-minute call How I can help

20+ years improving how teams deliver software, now helping organizations get past the AI pilot.

The AI factory I build with you

Automation is the accelerator: it pushes more work to the agents. Checks and reviews are the brakes, and good brakes are what let you go faster without drowning in low-quality pull requests. Here is each part, and what I bring to it.

Automation Accelerator starts agents on New ticketPull requestFailing pipelineNightly schedule I wire agents into your pipelines and schedules

Guides steer before it acts

Context & memoryI make your repos, docs and decisions readable for agents, so every task starts with the right knowledge
Conventions & pluginsI package your standards as plugins, one per goal: skills that teach the LLM your rules and conventions, plus the MCP connectors they need
Agent Claude CodeDevinCursorPiDeepSeek
LLM ClaudeGPTGeminiDeepSeekQwen I help choose the tools and roll them out to every team
Guardrails only for risky actions

Sensors Brakescheck after it acts

EvaluationI test your skills, plugins and rules like code: evals and compliance checks on every change
MonitoringI monitor the agents: which skills they use, the cost per successful task and how often a human must step in, so the data shows what to improve

Steering loop: every review becomes a rule, so the same comment is never written twice.

Foundations RepositoriesNx buildCI/CDTestsConventions I make them agent-ready

From ticket to production, at agent speed

Code is no longer the bottleneck: planning, review and maintenance are. Copying today's hand-offs onto agents only speeds up the old process, so each stage changes, and people keep the judgment.

Plan

Before
Requirements written by hand, without knowing the code.
The agent does
Turns an idea or a ticket into a clear intent, from the code and your context.
Human judgment
Product owner accepts the intent
Measured by
Time from idea to accepted intent

Spec

Before
Specs live in documents and drift from the code.
The agent does
Writes the spec as a versioned file in the repo, applying your security and compliance skills.
Human judgment
Tech lead approves risky changes
Measured by
Rework after the build starts

Design

Before
Static mockups handed off; every UI change waits for a developer.
The agent does
Builds live prototypes from your design system, so designers can ship UI changes directly.
Human judgment
Designers own the design system and the UX
Measured by
Time from idea to shipped UI

Build

Before
One developer, one task at a time.
The agent does
Plans the change, then codes it in parallel sessions.
Human judgment
Engineer approves the plan before any code
Measured by
First-pass merge rate

Test

Before
Tests written last, when there is time left.
The agent does
Writes and runs the tests until they pass; a second agent verifies.
Human judgment
People own the evals; CI decides
Measured by
Eval pass rate

Review

Before
Every line waits for a human; standards vary by reviewer.
The agent does
Reviews every pull request against the spec and your standards, and pushes fixes.
Human judgment
Intent and risky changes, not every line
Measured by
Review comments acted on

Release

Before
Release steps run by hand from a checklist.
The agent does
Prepares and runs the deployment, sandboxed.
Human judgment
A named person authorizes production
Measured by
Lead time to production

Maintain

Before
Incidents triaged by hand; upkeep loses to features.
The agent does
Triages incidents, keeps dependencies current and turns routine fixes into pull requests.
Human judgment
A human authorizes the fix
Measured by
Time to diagnosis, cost per team

Every session and incident feeds back as context, a new intent or a new eval, and the loop starts again. Git keeps the audit trail.

More than building: advice and advocacy

The tools change every quarter. Knowing where the market goes, and bringing people along, matters as much as the code.

Advise and challenge the strategy

An outside view on your AI plans, from someone who rolls them out.

  • Which tools, in which order
  • Build or buy, and where the risks are
  • Pushing vendors for enterprise features

Keep you ahead of the market

I test the tools myself and turn the noise into clear recommendations.

  • Claude Code, Devin, Cursor, Pi, DeepSeek
  • Harness engineering, skills, MCP, factories
  • Short, opinionated notes on what matters

Advocate for developers

Adoption is a people topic. I explain, demo and teach until it sticks.

  • Workshops and internal demos
  • Conference talks, in English and French
  • Articles on what works in practice
Nx Champion badge

Nx Champion. Speaker at React Brussels, Monorepo World, DevFest Nantes, This Is Learning Conf. 25 articles and 4 talks

How I work

The principles I bring to every engagement, learned from rolling out agents in real organizations.

  • Don’t outsource the thinking

    Agents amplify the design you did, or the lack of it. The architecture and the plan stay human decisions.

  • Review where the leverage is

    Research and plans get the closest human review. A wrong plan costs far more than a wrong line of code.

  • Brakes before accelerators

    Raise the quality bar with checks and review agents before adding more agents. Otherwise you only produce more rework.

  • Stay in the smart zone

    Small, focused contexts beat long conversations. Sub-agents control context; they are not job titles.

  • Measure, then decide

    Telemetry and evals decide which plugins to improve or retire, not opinions or vendor demos.

  • Pick one tool and get reps

    Adoption is a culture change led from the top, with depth on one tool rather than comparing five. Without it, juniors ship more with AI while seniors clean up after it.

20+ years, one thread

Every step has been about how teams build software better. AI is the next step, not a change of direction.

  1. 2006

    Developer

  2. 2013

    Lead developer & Scrum master

  3. 2017

    DX architect & Nx

  4. 2024

    AI in the SDLC

  5. 2026

    AI engineering factory

What people say

He is my go-to person for technical questions, not simply because he knows the technology, but because he is constantly exploring it hands-on. I regularly ground my own thinking in his insights before making decisions.

Read the full recommendation
Thomas Mauroy
Enterprise AI Lead at Swift

He reshaped how we think about and work with our tooling from the ground up. Jonathan also brought strong AI skills to the table, introducing AI carefully and intentionally, only where it genuinely made sense.

Read the full recommendation
Jose Badeau
Head of Technical Excellence at Caseware

He has the energy and passion, the vision, and technical expertise to drive a team, suggest roadmaps, and implement solutions at an amazing speed and quality.

Read the full recommendation
Miguel Perez Sanchis
Software Engineer at Julius Baer

Jonathan was entrusted with the challenge of leading our teams through the migration of a substantial legacy frontend codebase to a modern, state-of-the-art monorepo.

Read the full recommendation
Karel Frederix
Senior Software Engineer at Marigold

He sports a can-do mentality and manages to raise the quality bar across an Engineering department. Sound boarding with him does not get boring: never condescending, always enlightening.

Read the full recommendation
Igor Kalders
Engineering Manager at Marigold
All 18 recommendations

Organizations I've worked with

  • Swift
  • Julius Baer
  • Caseware
  • Entain
  • Marigold
  • Thalys
  • Sodexo
  • Thales
  • European Commission
  • Fednot
  • RTBF
  • Roularta
  • Edebex
  • I.R.I.S.

Latest writing

What I learn in real engineering organizations.

All articles and talks

Let's talk about your AI factory

A stalled pilot, agents that ignore your standards, no idea what the tools really bring? A 30-minute call is usually enough to see where I can help.

Book a 30-minute call or write to [email protected]