Context engineering is the real bottleneck for coding agents at scale

The system. LinkedIn built a context engineering layer for coding agents because out-of-the-box agents failed on its mature codebase. The company runs thousands of repos and services with custom databases, observability, feature flags, and experimentation systems. New engineers need a week-long bootcamp plus several weeks to become productive. Agents without access to that internal knowledge produced subpar code or hallucinated, so engineers stopped using them and went back to manual coding.

The mechanism. The core is an MCP server preinstalled on LinkedIn laptops, serving tools that wrap internal systems and playbooks that act as procedural memory. A playbook contains a name, description, and step-by-step instructions for a task like "write an Airflow pipeline" or "debug this service." Agents invoke playbooks as tools, so instructions enter the context window only when relevant.

Design rules. Playbooks are self-contained and composable. Each does one thing and can reference smaller playbooks, which gives progressive disclosure: an agent loads only the context needed for the current step and avoids context overload. This is essentially tribal knowledge turned into a navigable graph, with teams owning the playbooks for their workflows.

Scaling context. MCP performance degrades beyond roughly 30 exposed tools, so LinkedIn does not expose all tools directly. Instead three meta-tools let agents search the tool and playbook catalog by keyword and domain, fetch the schema for the chosen tool, then execute it. That is how 600 playbooks and thousands of tools stay usable without flooding the model.

Results. LinkedIn reports 8,000 daily users, including PMs, designers, and TPMs, and a 20% productivity gain with system reliability unchanged. The popular use cases are debugging and incident investigation, boilerplate code, migrations, pipeline management, and environment setup. The evidence is production experience, not a controlled benchmark, and the 20% figure is stated without methodology. No per-team breakdown or before-and-after comparison is given.

Maintenance. Playbooks go stale, so LinkedIn relies on agents to notice outdated instructions, improvise with code search and docs, and propose updates. Every tool and playbook goes through security review, and unused ones are aggressively deprecated. System reliability staying flat is the more meaningful signal. The lesson for engineering leaders is that internal context infrastructure, not model capability, determines whether coding agents deliver.