Coding & agentsGitHub

RAG, MCP and skills: ask which problem each one solves

GitHub’s September 18 podcast article explored debates about reading code, retrieval and agent tools. The useful follow-up is to separate the layers in an agent system.

A green library cabinet, copper bridge connector and lavender instruction scroll feed separate paths into one task station.
Conceptual illustration.

What changed

GitHub’s September 18 podcast article explored debates about reading code, retrieval and agent tools. The useful follow-up is to separate the layers in an agent system.

Why it matters

Retrieving relevant information, connecting a tool and supplying reusable workflow instructions are different jobs. They can coexist. A fashionable replacement narrative can distract from a simple question: which missing context or capability caused a failure? Read the actual tool contract and inspect what information reaches the model.

What to check next

Draw the path from user request to retrieved material, tool execution and final verification. Remove a layer only after testing that its job is still covered. Check permissions separately from retrieval quality. This explanation is Codorc’s analysis of the engineering questions raised, not a transcript of the podcast.

Event dates refer to the original announcement. Publication dates show when this article first appeared on Codorc.

Continue reading

All stories
Small green mechanical helpers move translucent code blocks along copper rails beside a manual review lever.
Coding & agentsOpenAI

Agent operations need a working stop button

OpenAI described Control Room on November 9 as a system for tracking and stopping model workloads across its research infrastructure. The useful design lesson is operational visibility.