Ajay Prakash describes why coding agents struggled with LinkedIn's internal frameworks, services and infrastructure despite their ability to generate code. Access to code search, documentation and other internal tools improved answers, but complex tasks still required scattered organizational knowledge, overloaded context windows and repeated rediscovery of prior findings.
LinkedIn's contextual agent playbooks expose task instructions through MCP alongside operational tools. An agent discovers a relevant playbook, retrieves its instructions and calls the tools needed to complete the work. Small, self-contained playbooks can be referenced by larger ones, making instructions reusable and allowing context to load progressively as a task requires it.
Ajay Prakash describes an improvement loop in which agents identify outdated instructions, discrepancies and missing information after using a playbook, then propose repository updates through pull requests. A local MCP server distributes central playbooks and tools to employee laptops, while repository-specific playbooks become available when an agent works in the corresponding codebase.
To avoid presenting an agent with every tool definition at once, the system exposes three meta tools: search, get schema and execute. Ajay Prakash reports adoption across engineering, product management and design, and argues that useful enterprise agents require reliable context and supporting infrastructure alongside capable models. An incident-response example includes seeking human confirmation before mitigation actions.
Watch on YouTube
