Show HN: A faster coding agent than Codex and Claude Code

**Bullet** is a coding agent introduced on Hacker News as a Show HN project that claims to run faster than **Codex** and **Claude Code**. The builders said their company was burning hours waiting on agent runs, and that the bottleneck was not the underlying models but the agent loops wrapped around them.

Those loops were slowed by unnecessary planning, oversized context, repeated searches, and serial tool calls. **Bullet** was built as a passion project aimed at cutting that overhead so developers can iterate faster without giving up quality.

On the product side, the agent automatically selects the model and reasoning level for each prompt. That is the main mechanism described so far: match model choice and reasoning depth to the task instead of running every turn through a fixed, heavy loop.

For engineers and builders, the pitch is practical rather than abstract. If most of the wait is orchestration—planning you did not need, context you did not ask for, search work done more than once, and tools called one after another—then tightening the loop can free wall-clock time even when the same class of models is still doing the hard work.

The competitive frame is explicit: **Bullet** is positioned against **Codex** and **Claude Code** on speed of the agent experience, not on a claim that it uses a superior base model. The differentiator is the loop design and per-prompt model and reasoning selection, set against agents that already set the bar for coding workflows.

What to watch next is whether automatic model and reasoning selection, plus removal of those loop inefficiencies, holds up outside the builders’ own workload: fewer stalled runs, less redundant search and tool chatter, and iteration that stays fast without quality slipping.