Mitchell Hashimoto builds a non-trivial Ghostty feature with Amp
Ghostty's creator walked through the agent sessions he used to build in-window, non-intrusive update notifications for the macOS app, including the planning, the dead ends and the cleanup.
The problem
Most AI coding demos are toy projects. This one shows an experienced maintainer shipping a real feature in a large production codebase with an agent, and where human judgment still mattered.
How they did it
- Start by writing a full plan interactively with the agent before any code.
- Hand the agent well-scoped fill-in-the-blank tasks, steering and reviewing as it goes.
- Finish with cleanup sessions, especially of the view model between UI and logic.
Results as reported by Mitchell Hashimoto
- The work took a total of 16 separate sessions totalling $15.98 in token spend on Amp.
- The author warns never to ship AI-written code without a thorough manual review.
These figures are as reported by Mitchell Hashimoto in Mitchell Hashimoto's blog ↗ (Oct 11, 2025). AgentGid did not measure or independently verify them.
This suits an experienced developer who already knows their codebase well enough to write a plan, scope tasks and catch bad output, since the human judgment is what keeps the sessions cheap. Watch the $15.98 figure: it's one author's spend on Amp, and Amp ranks 21st in its category with a Gid Score of 31, so expect different results elsewhere. Pi (Free, OSS) is a no-cost alternative, with only model provider costs.
The agent used here
Similar use cases
JustHTML: a Python HTML5 parser built with coding agents that passes html5lib tests
A developer spent his off-hours using Copilot agent mode with several models to build a dependency-free HTML5 parser, and wrote up every stage, including a full restart.
JustHTML is about 3,000 lines of Python with 8,500+ tests passing. (as reported by the author's blog)
Mid-Ohio Food Collective ships features in 24 to 36 hours with Claude Code
Mid-Ohio Food Collective, Ohio's largest food bank, also builds FreshTrak, the software that runs at some 2,500 pantries, soup kitchens, shelters and food banks. Its small enginee…
New features go from idea to production in 24 to 36 hours, down from two to three months, as reported by Anthropic.
IMDEX built a mining data platform in 8 months with Cursor
IMDEX, an Australian mining technology company, used Cursor to merge fragmented geological data products into one platform and to move a legacy Angular app to a React micro fronte…
The data platform landed in eight months; without Cursor it would have taken 20 additional engineers, two to three years and millions of dollars, as …