Eliver Sun

[Eliver Sun @eliver_sun]

@DurovPD

[Eliver Sun @eliver_sun]

GPT-5.6 in Codex doesn't just write code—it challenges the product. While building my local-first iOS Pokedex, it surfaced the truth: if the demo can't tell noodles from ice cream, it isn't useful. Then it helped redesign the recognition stack, UX, and tests around that truth.

[Eliver Sun @eliver_sun]

The more I use AI agents, the more I think the real gap isn’t whether they can produce an answer.

It’s whether they can reliably put the result where a human expects it to be.

Intelligence is the beginning. Reliability is the threshold.

[Eliver Sun @eliver_sun]

AGENTS.md gives the agent memory.

Implementation notes give the human leverage.

When the spec gets blurry, I don’t just want code. I want the trace:
what changed,
what was assumed,
what was skipped,
what needs review.

Autonomy only helps if the work is still inspectable.

[Eliver Sun @eliver_sun]

AI coding agents make one thing obvious:

many vague software tasks only worked because a human kept silently repairing the spec.

The model can write code, run tests, even debug.

But without a clear done, the agent mostly reveals that the task was never shaped.

[Eliver Sun @eliver_sun]

My current test for agent workflows:

If I delete the clever prompt, does the workspace still tell the agent what to do next?

If yes: rules, files, queues, review gates, fallback paths.
If no: I probably built a demo, not a system.