Senior Platform Engineer — Remote (Spain → EU)
Your enterprise customers shipping production loops on Mistral end-to-end — same MCP-native stack I run today, day one with me.
The threat nobody is pricing in
In 12 months, a single engineer with the right agent stack will out-iterate most product teams. The teams that survive aren't the biggest — they're the ones whose know-how is documented as the work happens, and whose infrastructure is agent-ready so the next pivot is a prompt instead of a rebuild. I help your platform team build both, in parallel with the shipping you already have to do.
Document the know-how
Decisions, runbooks, and tribal knowledge captured by the same agents that do the work — not a quarterly Confluence audit.
Make infra agent-ready
MCP-native internal tools. Next pivot is a prompt away, not a rebuild.
I already run this stack
Four generic crews end-to-end on hardware I bought myself. I know how it breaks before it breaks for you.
Speed is the moat
Your customers will pick the model that's cheapest plus the platform that ships features fastest around it. The first half of that you already own. The second half is platform work that compounds: tracing across La Plateforme + Le Chat enterprise, eval harnesses that run on every release, agent-loop primitives that customers can adopt without rewriting their stack. The wedge is whoever ships those weeks faster — the headcount on the other side of the bet doesn't matter if the loop is run by one engineer who already does this.
What your competition is doing
You compete on the model. You compete harder on the platform around it. Three peers are public about exactly that pivot — use them as the upper-bound reference for what platform work should look like in your 2026.
Anthropic
Open-sourced MCP — turned tool-use into a portable protocol the whole industry adopted.
MCP is now the substrate every serious agent platform speaks. Mistral's tool-use API speaks JSON-schema today; the gap to MCP-native is small but real, and it's the kind of platform decision that compounds for two years.
Source: anthropic.com/news
Hugging Face
Built the platform layer (Inference Endpoints, Spaces, Hub) into the moat — not the models.
HF's leverage is platform-on-top-of-everyone's-models. Same shape as the customer-platform work Mistral could ship around La Plateforme + Le Chat enterprise; the bottleneck is platform engineers who think in loops, not just inference.
Source: huggingface.co/blog
Cohere
Command-R + Toolkit — bet enterprise distribution on tool-use + RAG-as-a-platform, not raw model wins.
Cohere's moat isn't a single benchmark; it's the platform their customers use for tool-use + retrieval. That's the same opportunity Mistral has with La Plateforme — and the kind of work I'd ship rather than try to outscore Llama on a leaderboard.
Source: cohere.com/research
None of these win on model alone. The pattern holds for Mistral — the platform team is where the next 12 months of customer retention is fought.
Closest match — agent platform
I've already shipped the boring infra around interesting agents (loops, retries, evals, observability) — the work that makes a model platform deployable by customers who are not AI-native. That's the gap I'd close at Mistral.
Self-hosted on a 2-node cluster I bought myself: CrewAI crews, Hermes orchestrator, OpenHands autonomous coder, custom MCP servers, Langfuse traces, Postgres + Qdrant, the lot. Four generic crews (architecture, development, marketing, hr) pluggable into different "companies" — same primitives Mistral's enterprise customers will need to wire up around La Plateforme.
What I'd do in the first 90 days
First wedge: an MCP-native adapter on top of Mistral's tool-use API so customer agent loops can speak the same protocol they already use against Anthropic + everywhere else — a one-week compatibility win that closes the most-cited gap in community posts. Second wedge: a Langfuse-style trace + eval harness wired into La Plateforme so customers can defend "Mistral's model is enough" with cost + quality data, not vibes. Third wedge: pick the highest-friction enterprise customer onboarding flow and rebuild it as an agent that handles 70% end-to-end — same pattern I shipped at a fintech SaaS, daily failures to near-zero. The pattern, not the feature, is what compounds.
If after a short call it isn't the right fit, no pressure. Either way you get the analysis I already wrote about your stack.
Fidel Perez · Agent-platform engineer · 11+ yrs · MCP-native