Agent platform
Durable agent workflows on Temporal executing tens of millions of actions per month, with human review before every send and realtime streaming to the product.
Temporal · event-driven · streaming
LÉONARD HENRIQUEZ
I build AI agents that run in production, and the platforms underneath them.
I'm co-founder & CTO of Topo (YC W24), where our AI SDR went from $0 to $1M ARR in three months.
Coding since I was 12. Based in Paris.
- $1M ARR
- three months after launching our AI SDR
- 10M+
- agent actions executed per month
- 1B
- documents in our search index
- ×4
- merged PRs per month after moving to coding agents
WORK
Since 2022 I've been building Topo: AI agents for B2B outbound sales, used by 150+ customers including Air France and Aircall. As CTO I designed the platform, hired and led the team, and stayed its top committer with 1,205 merged PRs.
Agentic engineering
Redesigned how we ship around coding agents (Claude Code, Codex, Cursor): strict CI gates, shift-left feedback, auto-fix loops, background agents and automated incident triage.
Merged PRs/month 80 → 318 · bugs per 100 PRs −63% · CI failures 15.7% → 2.4%
LLM systems
Agents for account discovery, enrichment, personalized outreach, reply handling and in-product chat. Context engineering over CRM data and call transcripts, with LLM evals to measure and improve quality.
context engineering · RAG · tool use · evals
Playbooks & MCP
A composable, trigger-driven workflow engine (signals → agent steps → sequences) with versioning, test runs and rollback. An MCP server with 170+ tools lets customers run Topo from their own agents.
workflow engine · MCP
Search at scale
The data layer: a 1B-document Elasticsearch index with hybrid search and ranking, migrated from Quickwit and fed by durable enrichment pipelines.
Elasticsearch · hybrid ranking · PostgreSQL
Enterprise readiness
Made Topo ready for customers like Air France: CRM integrations, SAML SSO, security reviews, and SOC 2 Type 2, which I led.
SOC 2 Type 2 · SSO · CRM integrations
PRINCIPLES
Business school and product taught me to tie every architecture decision to a business outcome. These are the lessons from shipping agents to real customers that held up.
- 01
Agents are distributed systems
An agent run is a long-lived workflow that calls slow models and flaky APIs. It needs what any distributed system needs: durable execution, retries, idempotent steps, replay. That is what lets Topo run tens of millions of actions a month.
- 02
Context beats prompts
Most of our quality gains came from what the model sees, not from how we ask: CRM history, call transcripts, the right tools at the right step. Evals tell us whether a change actually helped.
- 03
Keep humans where mistakes are expensive
Every email Topo sends is reviewed by a person first. Autonomy should be earned one action at a time, as the evals show the agent is ready.
- 04
Coding agents need feedback, not freedom
Strict CI gates and fast feedback let agents catch and fix their own mistakes. That took us from 80 to 318 merged PRs a month with 63% fewer bugs per PR.
TALKS
Talks, juries and recognition.
- 2024Talk: LangGraph in production
- 2024Judge, Mistral AI hackathon, San Francisco
- Judge at AI hackathons in Paris
- Invited to the Élysée by President Macron to help prepare the AI Action Summit
EARLIER
From product management to running engineering.
- 2019–22Head of Engineering at Flatlooker (now Manda). Started as product manager, then ran engineering: ~10 developers, recruiting and training.
- 2015–19MSc Data Analytics & AI at EDHEC. Product manager at BTU Protocol; data science intern at Jalgos, where I wrote a neural network from scratch in C++.