Top 200 ReposEXPLORER
Browse agents, categories and pages
Back to results
Reported by sourceThe source says this happened. Top 200 Repos has not reproduced it.

Ship a non-trivial app feature over many short agent sessions

Mitchell Hashimoto built Ghostty's unobtrusive macOS update notifications with the Amp agent over 16 sessions, rewriting and reviewing parts by hand before merging.

The problem

How it works

He planned the feature with the agent and its read-only Oracle subagent, then used it for a UI prototype, filling in scaffolded functions, refactors, documentation and update-flow simulations, stopping or discarding work that went wrong. He personally researched a titlebar bug, switched to an overlay design and restructured the view model before letting the agent continue.

Results and limitations

He reports about 8 hours over three days and $15.98 in token spend across 16 sessions; the agent was strong on prototypes, well-scoped tasks and cleanup, but repeatedly failed on a layout bug and produced a backend he threw away.

Before you try it

Substantial human coding and review shaped the final code, and the time figure is his own rough estimate. The feature first reached only nightly users, and Amp is a hosted tool without a repository in this directory.

Tools and repositories

No ranked repository is involved. This use case still stands on its own source.

Tools
Amp · Sparkle

Try it yourself: Plan and build a feature in small reviewed sessions

Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].

Preview
We are adding [Feature] to [Project]. In this first session, do not write code. Read the relevant parts of [Relevant code area] and produce a plan of small, independently testable steps, with open questions and risks. After I approve, each later session handles one step only: implement it, build, run [Build and test command], and summarize what changed. If an approach fails twice, stop and explain rather than trying more variations. End every session by listing anything obvious you may have missed. Do not commit, push or touch unrelated files; I will review and commit selectively.

Understand every change before keeping it, and discard agent code you cannot explain.

Sources

  1. Vibing a Non-Trivial Ghostty Feature (opens in a new tab)

    First-person account · mitchellh.com · Published 11 Oct 2025 · Source checked 9 Nov 2026, 22:21 UTC

    What this source supports
    • Amp with its Oracle subagent used to build update notifications
    • 16 sessions, $15.98 in tokens, about 8 hours over three days
    • Agent failed on a titlebar layout bug; backend draft discarded
    • Author did manual restructuring and final review
Report a problem

Reported by source

Build an app with Replit Agent while keeping production data out of reach

Replit Agent builds and browser-tests apps from a description. A widely reported 2025 incident, where it deleted a live database during a code freeze, shows why separate dev and prod data matters.

docs.replit.com
Reported by source

Explore and change a large codebase from the terminal with Gemini CLI

Gemini CLI is Google’s open-source terminal agent. Simon Willison found it strong on large codebases in 2025; in 2026 Google moved individual users to Antigravity CLI.

github.comgoogle-gemini/gemini-cli
Reported by source

Hand coding tasks to Cursor cloud agents that open pull requests

Start Cursor agents from the editor, web, Slack or a GitHub comment; each works in its own cloud VM on a branch and returns a pull request with screenshots, videos and logs.

cursor.com
Reported by source

Rebuild a side project in OpenCode to test provider lock-in

Matthew Brunelle rebuilt a YouTube Music to OpenSubsonic shim with OpenCode and Kimi K2.6 to compare it with an earlier Claude Code and Opus 4.6 attempt.

blog.matthewbrunelle.com · 30 Apr 2026anomalyco/opencode
Planned, not done

Run coding agents from one self-hosted control centre with OpenHands

OpenHands Agent Canvas runs the OpenHands agent, or Claude Code, Codex, Gemini CLI and other ACP agents, on your machine or in Docker sandboxes, and can trigger them from schedules or webhooks.

github.comOpenHands/OpenHands
Reported by source

Apply the context engineering lessons from Manus to a long-running agent loop

The Manus team describes the context rules it settled on after rebuilding its agent framework four times: keep the prompt prefix cacheable, mask tools instead of removing them, and use files as memory.

manus.im · 18 Jul 2025