Becoming your team's AI ambassador without asking for it
TL;DR
The tool wasn't the problem, the model was. Kilo Code + Mistral looked perfect on paper. In practice, Mistral nearly killed adoption. Pivot to OVH AI Gateway + Gemini. It works now.
There was no official appointment. At some point, I was the only one on the team who’d been using coding agents daily for a few months, and word got around. The questions started: “got an opinion on Cursor?”, “shouldn’t we try that?”, “could you give us a presentation?”
That’s how you become an AI ambassador in a tech team — by default, because you’re visible on the topic.
Selecting the tool
The core constraint: the team is heterogeneous. Some devs on VS Code, others on JetBrains, and me in terminal. A tool that only exists in one ecosystem creates second-class citizens on the team from day one.
The criteria we set collectively:
- Compatible with VS Code and JetBrains and CLI — no compromises
- Multi-model: no lock-in to a single provider
- Open-source or at minimum auditable code — no black box
- No fixed per-dev fees — pay-per-use billing
Kilo Code checked every box. Native VS Code, JetBrains plugin available, CLI for terminal profiles, Apache 2.0, multi-provider by design. The decision was quick — it was objectively the best choice given our criteria.
Choosing Mistral — a good idea on paper
For the model, the initial proposal was Mistral via their free API. Two strong arguments: the free tier for a first test without committing budget, and sovereignty — Mistral is French, European infrastructure, GDPR by design. For a team working on client projects with sensitive data, that’s a real argument, not a symbolic one.
On paper, it was the perfect choice. Free API + open-source tool + sovereignty = frictionless adoption.
What actually happened
The team installed Kilo Code. Configured the Mistral API. Started working.
And it stalled — not because of Kilo Code, but because of Mistral. On concrete coding tasks in our Symfony/TypeScript stack, the model produced results below the expectations set by demos and benchmarks. Functional code but not idiomatic. Suggestions that didn’t respect project conventions. Refactoring instructions followed incompletely.
This isn’t a general criticism of Mistral — it’s a good model for many use cases. But for agentic coding on a specific PHP/TypeScript stack, the comparison with Claude Sonnet or Gemini was unfavorable.
The problem: the first few days with a tool shape perception. Devs who could have become regular users associated Kilo Code with a disappointing experience. Some simply stopped trying.
The tool wasn’t at fault. The model was. But for someone making their first contact with coding agents, that distinction isn’t obvious.
The pivot
The decision was to change the model without changing the tool. Kilo Code stayed — the configuration, the custom commands we’d started building, the nascent habits. We just swapped the engine.
OVH AI Gateway became the hub. It’s an OpenAI API-compatible proxy that provides access to multiple providers through a unified interface, hosted in Europe, billed in euros. The advantage for the team: a single API key to manage collectively, one place to see consumption, and the ability to switch models without redistributing credentials to every dev.
Gemini 3 has been the default model since. On our stack, the quality difference compared to Mistral is notable — the code is more idiomatic, complex instructions are better followed, and the context window is comfortable for loading Symfony files without fine-grained slicing.
The results changed. Adoption too — devs who had dropped off gave the tool another chance with the new model, and this time the experience met expectations.
What changed since
The team moved to the Anthropic API directly — no longer through OVH AI Gateway. The motivation: Claude Sonnet’s quality on architecture and complex refactoring tasks, and direct access to the Anthropic ecosystem. As of mid-2026, that’s still the setup in place: Kilo Code VS Code + Anthropic API for the team, Gemini 3 available for day-to-day work.
On the personal side, the trajectory kept moving: I went back to Claude Code on a Max subscription — not to replace the team setup, but because the Anthropic ecosystem (agents, hooks, persistent memory) is worth the individual investment. I document this U-turn in Anthropic, the Apple of AI.
What this taught me about adoption
Choosing the tool is the easy part. You can evaluate objective criteria, make comparisons, rationalize a decision. That’s engineering work.
Managing adoption is different. It’s the first few days that matter — not the features, not the benchmarks. If the first experience disappoints, regardless of the tool’s intrinsic quality, most people won’t invest in understanding why it didn’t work.
On the model specifically: the free tier is a secondary criterion for initial tests. What matters is that the results are good enough that people who aren’t used to agents yet want to come back the next day. If the free model produces decent results, it’s the right choice. If it produces results that frustrate already-skeptical devs, the cost of failed adoption far exceeds that of a paid API key.