Agentic AI
LinkedIn's Playbook Strategy: How AI Agents Learn to Navigate Complex Enterprise Systems
LinkedIn engineers built procedural memory structures called "playbooks" using the Model Context Protocol to equip AI coding agents with institutional knowledge and tackle complex tasks reliably.
InfoQ
Engineering teams at major tech companies face a unique challenge: off-the-shelf AI coding agents perform well on generic tasks but falter when deployed against massive, interconnected codebases built on internal frameworks and custom infrastructure. At LinkedIn, engineers attempted to use coding agents for incident response and development workflows, but the approach largely failed because the agents lacked organizational context. This experience led LinkedIn to design an architecture that fundamentally changes how agents access and retain institutional knowledge.
The Vibe Coding Problem at Scale#
When early AI coding assistants gained the ability to not just suggest code but actually execute terminal commands and edit files directly—a trend some call "vibe coding"—teams at LinkedIn eagerly adopted the technology. However, the premise quickly broke down. LinkedIn's codebase spans thousands of repositories and microservices, all built atop layers of custom databases, tracking systems, and observability infrastructure. New engineers spend a week in onboarding just to familiarize themselves with these internal systems; no LLM trained on public open-source code could navigate such complexity autonomously. Agents either generated low-quality code or hallucinated solutions entirely, forcing engineers to babysit them with constant guidance—the opposite of productivity.
Building a Context Layer with MCP#
The turning point came when Anthropic released the Model Context Protocol (MCP) as an open standard for connecting tools to agents. LinkedIn wrapped its sophisticated internal code search engine—a system that allows engineers to query across 1,000+ repositories—as an MCP tool. This single addition was transformative: agents could now learn how code actually looked at LinkedIn by querying real examples rather than relying solely on pre-training. The team progressively added more MCP tools—document and wiki search, feature flags, task management, and data platform access—each one expanding the context agents could draw from.
Yet complexity remained. Even with richer information, agents struggled with end-to-end task execution on anything moderately complex. They faced three fundamental barriers: scattered tribal knowledge (compilation steps, testing commands, installation procedures buried in Slack threads and senior engineers' heads), context overload (ballooning tool outputs that consumed precious LLM context windows and caused agents to forget what they were doing), and the lack of persistent memory (agents had to rediscover the same solutions repeatedly, burning tokens and time).
