Valentin Bonnet

San Francisco, CA  ·  github.com/vbonnet  ·  linkedin.com/in/vbonnet

Staff-level software engineer with 7+ years at Google, YouTube, and Verily (Alphabet). Current focus: AI agent infrastructure — meta-harnesses, MCP tooling, context architecture, and testing frameworks that give non-deterministic LLM output measurable correctness guarantees. Active open source author in the AI agent engineering space.

Open Source

dear-agent — github.com/vbonnet/dear-agent

A pluggable meta-harness for autonomous AI coding agents, formed by merging two earlier projects (a harness layer and a memory/knowledge layer) into a single product.

  • Harness-agnostic adapter interface unifying Claude Code, Gemini CLI, Codex CLI, and OpenCode behind one session lifecycle. Copy-on-write workspace isolation (OverlayFS on Linux, APFS on macOS), state-aware async message routing, and advisory file reservations for safe parallel agent workflows.
  • DEAR protocol for research-loss prevention: enforces commit / push / human-readable branch / completion-notification gates before a session can be archived, eliminating the “stranded work on UUID branches” failure mode that plagues long-running agent sessions.
  • Persistent cue-based memory so agents retain context across conversations and sessions, with a metacontext Go service backing 3-tier (filter → rank → load) retrieval at <1ms p50.
  • Phased SDLC workflow engine (Wayfinder) with LLM-as-judge validation gates and a plugin architecture for MCP connectors and tool integrations. Go, Apache 2.0.

Skills

Go, Python, SQL. LLM integration, RAG pipelines, MCP (Model Context Protocol), multi-agent orchestration, non-deterministic testing, context window management, agent safety and correctness. Bazel, Kubernetes, distributed systems, CI/CD, developer platform engineering. gRPC, Protobuf, GCP, SQLite.

Work Experience

Senior Software Engineer, Verily (Google/Alphabet)

Tech Lead, Developer Platform & Velocity  ·  2021 – Present (parental leave through May 2026)

Verily is a precision health company under Alphabet (~1,000 engineers). Developer Platform owns the tooling, SDKs, and shared services the full engineering org builds on.

  • ViDA platform: Built MCP-based developer tooling and AI agent testing infrastructure adopted across the engineering org. Cut integration test setup time for new product teams from days to hours by abstracting environment provisioning and dependency injection.
  • Agent evaluation harnesses: Designed test frameworks for non-deterministic LLM output at scale. Product teams can now ship agentic features with measurable correctness guarantees rather than relying on manual spot-checks, reducing post-launch rework cycles.
  • Org-wide developer productivity: Owns internal SDKs and shared platform services used by all engineering teams. Drives adoption through developer experience design, not mandates.
  • Cross-team technical leadership: Sets cross-team technical direction, reviews architecture proposals, and aligns with product and infra leads on roadmap priorities.

Software Engineer, Google — YouTube

Backend Infrastructure  ·  2019 – 2021

  • Worked on backend infrastructure supporting YouTube at global scale (100M+ DAU).

Software Engineer, Google

Developer Tools & Build Infrastructure  ·  ~2016 – 2019

  • Built deep expertise in Bazel at Google scale: 100K+ generated build targets, reverse-dependency analysis tooling, and large-scale automated refactoring.
  • Contributed to developer platform and build infrastructure used across Google engineering.

Education

McGill University — Montréal, QC
B.A. in Computer Science  ·  2009 – 2013