Build Daily
Tinley Park · October 13, 2026My Coding Agent Took 19 Seconds to Say Anything
The video documents a coding agent running on an M5 Ultra Mac Studio that took 19 seconds before producing any output, and explores what the M5 Ultra requires from the user before it can truly perform. It is part of a build-in-public series on setting up local AI coding infrastructure on Apple Silicon hardware.
The outline
Every point the video makes, in the order it makes them.
- 01
Coding Agent Latency
The creator reports that their AI coding agent took a full 19 seconds before generating any response, highlighting a performance bottleneck.
- 02
M5 Ultra Requirements
The M5 Ultra needs a specific input or configuration from the user before it can pull ahead in performance, which is the central finding of the video.
- 03
Hardware Build
The creator shares gear used in the setup, including a cluster rack, a 40Gbps T4 enclosure, an NVMe SSD, and a 15-inch magnetic display.
- 04
Local LLMs on Apple Silicon
The video sits within a broader series on running large language models and AI agents locally on Apple Silicon hardware, including Mac Mini clusters and Neural Engine workloads.
- 05
Developer Productivity & AI Coding
The content is part of the creator's developer productivity and AI-for-coding playlists, focused on building practical local AI tooling for software development.
- 06
Build-in-Public Context
The video is framed as a transparent, step-by-step public build log, inviting viewers to follow the creator's iterative hardware and software setup process.
A coding agent that answers in 0.12 seconds on turn one can sit silent for 19 seconds by turn twenty-five, because every turn forces a reread of the full conversation on a single GPU. For a solo builder shipping agent-based features, the latency isn't a fixed cost—it multiplies with each interaction, and caching won't flatten the curve. The design question shifts from how much context to allow to how to bound it before the user waits.