LinkedIn caps 1,300 agent tools behind just 3 meta-tools

한국어로 읽기

LinkedIn's coding agents don't hold thousands of tools and playbooks in context at once — they search, fetch a schema, and execute, because MCP degrades past 30 or 40 tools. The talk traces how LinkedIn went from agents that hallucinated on internal code to a system handling an on-call incident end to end in minutes.

In their words

So the engineers had to prompt these agents manually um to do the right thing which used to take more time than the manual coding itself. So a lot of people a lot of engineers went back to manual coding.4:18

we cannot scale it beyond 30 or 40 tools without degrading the uh context or degrading the performance of the system. So what we do is instead of uh surfacing all of these playbooks and tools through MCP we replace them with three meta tools.17:04

So we have over,300 uh tools and over 600 uh playbooks and it's not not just engineering right.18:16

Disclosure · Prakash is a LinkedIn engineer presenting LinkedIn's internal system, which functions as a promotion of the company's engineering work.

One thing to add — One thing to add — the talk doesn't quantify error rates or time saved beyond the anecdotal "hours to minutes" claim, so the 8,000 daily users and 1,300/600 tool-playbook counts are the only hard adoption numbers given. It's worth watching whether the self-updating playbook loop (agents editing playbooks via PR) actually holds up against drift, since that's asserted rather than measured in the talk.

One thing to try tonight
Try splitting one of your own long AI-agent system prompts or docs into small, single-task "playbook" files referenced from a top-level index, then see if your agent picks the right one on a test query instead of reading everything at once.

Written from a single source read through its auto-captions; the original is what counts. Interpretation lives only in "One thing to add". Quotes and slides are quoted for reporting and criticism, with the source and timestamp shown. If you are the speaker or rights holder and would rather we took this down, we will — support@juniverse.app.