Solo dev ships native Mac app on multi-agent workflow: 5 lessons from Claude Code + reviewer agent
Moist_Tonight_3997 · reddit · 2026-10-04
A solo developer shipped v3.0 of Buffer, a native Swift/AppKit clipboard manager for Mac (MIT, fully local, 430+ stars), built entirely on a multi-agent workflow: Claude Code as the driver and a second model (opencode) as a dedicated reviewer on the same diffs. Five lessons, no hype:
- Agents execute specs, they don't write them. "Make zooming work" produced a physics-model monster; an explicit state machine (scale + offset, double-click to fit/reset, ⌘+/⌘- steps) landed in one pass. The bottleneck was the spec, never the model.
- The reviewer agent is the highest-leverage role. A driving agent defends its own diff; a fresh-context reviewer instructed to argue against it catches real edge cases and dead branches, in parallel, for minutes of cost.
- Native UI testing is still human — gesture feel (rubber-banding, anchor drift) can't be agent-verified.
- Community specs beat agent specs — users voted on whether Esc saves or reverts; once decided, execution was flawless.
- External PRs improve with agent review — the reviewer analyzed the first outside contribution's diff and wrote tests before merge.
Repo: github.com/samirpatil2000/Buffer
More from coding & agent
- 8 Emerging Standards for AI Agents You Should Know — bibryam · 2026-10-04
- Developer Guesses Open Coding Models Are Less Than Six Months Behind Frontier — heyneighbor · 2026-10-04
- Harness vs. Loop Engineering: The Two Distinct Problems in Building Reliable AI Agents — goyalshaliniuk · 2026-10-04
- ffmpeg-skill turns coding agents into video editors: 42 local FFmpeg tools, auto-verification — udmrzn · 2026-10-04
- Strata calibrate nearly tripled decode speed: 256K context on a 16GB GPU — MoonsvnLyn · 2026-10-04
- One prompt freed 39.7 GB: Claude Code + ccmd MCP safely cleans dev caches — julsimon · 2026-10-04