Platform Engineering for AI: Scaling Agents and MCP at LinkedIn

Moving Beyond Siloed Proof of Concepts

LinkedIn is enabling AI agents at enterprise scale by recognizing them as a new execution model—comparable to microservices—requiring shared, production-grade infrastructure. Karthik Ramgopal and Prince Valluri of LinkedIn detail how platform teams are orchestrating secure, multi-agentic systems and reducing toil for developers.

Why This Matters

Many organizations experiment with AI in isolated silos, leading to inconsistent infrastructure, duplicated effort, and difficulty scaling beyond initial proofs of concept. The cost of fragmented AI initiatives and duplicated infrastructure can quickly become substantial, hindering widespread adoption and ROI.

Key Insights

• LinkedIn has thousands of developers actively using AI-powered tools daily.
• Foreground agents assist developers directly in the IDE, while background agents autonomously handle repetitive tasks like code migrations.
• MCP (Model Context Protocol) provides a vendor-neutral standard for tool calling, enabling consistent agent interactions across LinkedIn’s ecosystem.
• LinkedIn uses RAG (Retrieval-Augmented Generation) and historical PR data to provide agents with crucial context, improving accuracy and relevance.
• Agents propose code changes via Pull Requests, maintaining human review and control throughout the process.

Working Example

(No code provided in context)

Practical Applications

• Use Case: LinkedIn utilizes background agents to automate dependency upgrades across its codebase, reducing manual effort and ensuring consistency.
• Pitfall: Overly restrictive agent controls can hinder developer autonomy and limit the potential benefits of AI-assisted workflows.

References:

• https://www.infoq.com/podcasts/platform-engineering-scaling-agents/