Sign InOpen Brain
AI EngineerVideoSource Linked

500 Skills, Zero Fine-Tuning: LinkedIn's Playbook for AI Agents — Ajay Prakash, LinkedIn

LinkedIn scales a large internal agent catalog through search, schema lookup, and execution rather than exposing every tool at once. Its playbooks add task-specific operating knowledge.

AI Engineer · Sep 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

LinkedIn’s internal system serves more than **1,300 tools** and **600 playbooks** to coding agents. Because performance reportedly degrades beyond 30–40 exposed MCP tools, the catalog sits behind **three meta-tools**: search, schema lookup, and execution.

Practical Implication

Use progressive discovery when your tool surface grows. Keep playbooks narrow and composable, load their instructions only when relevant, and let agents propose reviewed updates when real work reveals stale or missing guidance.

Agent-Ready Context
LinkedIn’s internal system serves more than **1,300 tools** and **600 playbooks** to coding agents. Because performance reportedly degrades beyond 30–40 exposed MCP tools, the catalog sits behind **three meta-tools**: search, schema lookup, and execution.

Use progressive discovery when your tool surface grows. Keep playbooks narrow and composable, load their instructions only when relevant, and let agents propose reviewed updates when real work reveals stale or missing guidance.

The reported scale includes **8,000 daily users** inside LinkedIn’s centrally managed environment. Its tool-limit observation is not presented as a controlled benchmark, and smaller teams may not need the authentication, telemetry, and distribution infrastructure described.
Connected Context · Feed7 Judgment

This turns progressive discovery from a library-design preference into an operating pattern demonstrated at unusually large internal scale: thousands of tools and hundreds of playbooks remain behind three meta-tools. It confirms narrow, composable, selectively loaded skills, while adding reviewed self-updates and central distribution; the observed tool ceiling and enterprise infrastructure should not be treated as universal requirements.

Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSetLinkedIn’s three-meta-tool catalog is a concrete large-scale implementation of FactSet’s claim that routing descriptions and progressive loading become product requirements as skill libraries grow.We Vetted 2000 AI Skills Before They Reached Developers — Lucas Palma, NubankA centrally distributed catalog of reviewed playbooks makes Nubank’s supply-chain controls directly consequential: discovery and scale also expand the need for scanning, gating, ownership, and reevaluation.Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI LabThe reviewed playbook-update loop operationalizes the shared, continuously maintained repository practices proposed for team-wide coding-agent adoption.How I automate my own job at Hugging Face using agents — Niels Rogge, Hugging FaceThe Hugging Face case provides a useful scale contrast: one bounded workflow can succeed with one CLI and one skill, whereas LinkedIn’s meta-tool layer addresses catalog complexity that smaller deployments may not have.
Context Map
agentcoding#harness-engineering#skills#tool-use
Uncertainty
The reported scale includes **8,000 daily users** inside LinkedIn’s centrally managed environment. Its tool-limit observation is not presented as a controlled benchmark, and smaller teams may not need the authentication, telemetry, and distribution infrastructure described.