Summary
Ajay Prakash, software engineer at LinkedIn, presents at AI Engineer how the company solved a fundamental problem: coding agents like Claude Code, Cursor, and GitHub Copilot are trained on public code and have no knowledge of LinkedIn’s internal infrastructure — over 1,000 repositories, proprietary databases, custom experimentation platforms, and configuration management systems that new engineers spend a full week learning in boot camp.
The solution, called Contextual Agent Playbooks, goes beyond providing MCP tool access. LinkedIn found that even agents with broad tool connectivity still hallucinated, got stuck, or produced untrustworthy code because critical tribal knowledge was scattered across wikis, Slack, and outdated docs. Three compounding failures drove this: agents lacked step-by-step procedural guidance, tool outputs progressively consumed context windows causing information loss, and agents had no durable memory between sessions.
Playbooks solve this by delivering instructions as MCP tool outputs — structured context that agents invoke on demand, just like any other tool. The system now provides 500+ skills with zero fine-tuning required. A flagship use case is on-call incident response: an agent fetches runbooks, queries logs and metrics, identifies root causes, creates a fix PR, and updates the incident management system in minutes rather than hours. Prakash walks through the architecture evolution from 2025 through production deployment, offering concrete guidance for enterprise engineering teams trying to make coding agents reliable at scale.
📺 Source: AI Engineer · Published September 09, 2026
🏷️ Format: Workflow Case Study
