Ian Thomas walks through Meta’s “Assess and Grow” maturity model, the community‑first rollout of AI‑for‑productivity tools, and concrete wins such as 90 % test coverage with AI‑generated diffs. He shares lessons on trust, tooling integration, and how teams can start their own AI‑native journey.

AI‑Native Engineering at Meta’s Reality Labs

In a June 25 2026 webinar, Ian Thomas (Software Engineer, Meta) presented a case study of how his Horizon Experiences team turned AI‑assisted tooling into a measurable productivity engine. The talk covered the Assess and Grow framework, community building, real‑world outcomes, and practical take‑aways for any organization looking to embed AI into its development workflow.

What’s New: The “Assess and Grow” Maturity Model

Meta’s engineering excellence program already focused on three pillars – implementation quality, engineering productivity, and production excellence. In August 2025 Ian introduced a six‑dimension maturity model that maps a team’s AI adoption from SIT (no AI usage) to LEAP (AI native). The dimensions are:

Workflow Integration – from “tool known” to “AI baked into daily pipelines.”
Prompt Skill & Sharing – from ad‑hoc prompts to reusable prompt libraries.
Trust & Quality – from frequent hallucinations to calibrated risk‑aware agents.
Observability & Feedback – from manual logs to AI‑driven alert triage.
Collaboration – from isolated experiments to community‑wide retrospectives.
Governance – from informal usage to policy‑backed guardrails.

Each dimension has five levels, and teams assess themselves in a retro‑style workshop (dot‑voting, discussion, and action‑item capture). The model is deliberately tool‑agnostic; it works with internal platforms like Devmate, RACER, and the orchestration layer Confucius as well as third‑party LLMs such as Claude or Gemini.

“The value is in the conversation, not the score,” Ian emphasized, warning against Goodhart’s Law.

Developer Experience: Building the Community and the Toolchain

1. Start Small, Iterate Fast

The initial pilot ran in a tightly scoped “1P Spaces” team. By limiting scope, the group could experiment, surface failures, and iterate without exposing the whole org to risk. Early adopters used Devmate (Meta’s VS Code partner) for assisted coding and RACER for risk‑aware test generation.

2. Grow the Community (AI4P)

Month 0‑2: 10 engineers formed a private Slack channel.
Month 3‑5: Organic word‑of‑mouth grew the group to 100 members.
Month 6‑9: Formal “lean‑coffee” sessions, office‑hours, and a two‑day training program pushed membership past 400.

The community kept the culture safe for failure, celebrated both wins and learning moments, and produced a shared process library of prompts, evaluation scripts, and best‑practice checklists.

3. Structured Learning

A two‑day curriculum split into:

Day 1 – Mindset & Foundations: AI‑native concepts, prompt engineering, risk awareness.
Day 2 – Hands‑On Tools: Live labs with Devmate, RACER, and building custom agents on Confucius.

Post‑training surveys showed a 30 % increase in confidence when using AI‑generated code suggestions.

User Impact: Concrete Wins and Measurable Gains

Area	Before AI	After AI (7 months)	Impact
Test coverage	70 % average	90 % on target repo (3 h effort)	60 % reduction in manual test authoring time
Diff size	Avg 1 k LOC	Avg 2.5 k LOC (larger but vetted)	40 % faster feature delivery, offset by improved review tooling
Weekly active tool users	35 %	80 %	Higher adoption without mandatory mandates
Incident rate (post‑merge)	1.2 %	0.7 %	Early risk‑aware agent (DRS) flagged high‑risk changes before merge

Highlight: Automated Test‑Coverage Sprint

A team used RACER to:

Identify the 15 % of files with the highest risk of uncovered bugs.
Generate test stubs automatically.
Parallelize diff creation across multiple agents.

What would have taken ≈ 20 h of manual work was completed in ≈ 3 h of human oversight, yielding 90 % coverage and 60 merged diffs.

Highlight: Voice‑First Development

Using Wispr Flow, engineers dictated code at ~140 wpm versus ~90 wpm typing. The tool fed spoken snippets into Devmate, cutting the edit‑refine loop by roughly 25 % for documentation and test‑driven development.

Lessons Learned (The Hard Parts)

Trust & Quality: Early models produced “slop” and hallucinations. The team mitigated this by layering prompt‑based rules (Claude skills) on top of the agents and by requiring a human‑in‑the‑loop review for any diff larger than 500 LOC.
Measuring ROI: Weekly active users proved too shallow. Meta shifted to metrics such as time saved per workflow and defect reduction per AI‑generated change.
Code Review Overload: Larger AI‑generated diffs strained reviewers. The solution was a pre‑review agent that flags high‑risk sections, letting reviewers focus on the most critical parts.
Accountability: When an AI writes code, the original author remains accountable. Clear ownership policies and audit logs in Confucius helped maintain traceability.
Tool Evolution: The tooling landscape changed dramatically within months (Claude 3 → Gemini 1.5). Continuous re‑evaluation of prompts and model versions is now a standing agenda item for the AI4P community.

Playbook for Your Team

Create a Safe Experiment Zone – Start with a single, well‑bounded team.
Run an “Assess and Grow” Workshop – Use the six dimensions to surface gaps.
Pick One Bounded Problem – E.g., test‑coverage gaps, code‑style linting, or API‑usage discovery.
Build a Minimal Agent on Confucius or an internal LLM; keep the prompt simple and iterate.
Add Human Review – Deploy a rule‑based guardrail that only escalates high‑risk diffs.
Scale the Community – Share successes, host lean‑coffee sessions, and publish a living prompt library.
Measure Outcome‑Focused KPIs – Time‑to‑ship, defect rate, and engineer satisfaction rather than raw usage counts.

Looking Ahead

The AI‑native journey is far from finished. Upcoming focus areas include:

Extending the maturity model to embedded and mobile teams.
Building domain‑specific MCP agents that surface runtime state for VR worlds.
Refining anti‑slop rules with automated post‑merge analysis.
Exploring continuous‑learning pipelines where model updates are validated against internal test suites before rollout.

“When we offload the heavy lifting to AI, engineers get space to explore, innovate, and ship the ideas that truly move the product forward.” – Ian Thomas

AI‑Native Engineering at Meta’s Reality Labs: From Manual Toil to Autonomous Productivity