Launch HN: Bullet (YC S26) – A Faster Coding Agent
codewithbullet.com adi1
Bullet says it built a faster coding agent after getting sick of waiting on Claude Code and Codex. It claims fewer round trips, lower cost, and 119-second SWE-bench runs.
Based on reporting by codewithbullet.com, adi1 — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Bullet is the latest “we built the thing we wanted” startup story, except this one started with six pivots, a dorm room, and a growing hatred of waiting around for coding agents to finish. Founders Adi and Alex say they came out of AppLovin and Citadel thinking they were on a straight path to startup glory. Instead they tried an AI hedge fund, a browser-use agent, synthetic financial data, a mobile IDE, and other ideas they describe as terrible, useless, or both.
The real trigger was much simpler: Claude Code and Codex were eating their time. Bullet says the founders were spending hours waiting on those tools, then decided to build something that attacked the part that hurt most. The result is a coding agent that leans hard on faster search, tighter context handling, and fewer back-and-forths. It routes tasks to different models, uses targeted code and context search rather than stuffing the whole repo into memory, keeps tool output bounded, drops stale screenshots, and avoids rereading files when it can help it.
The company also says it tries to make each turn count. Independent investigations happen in parallel, while edits and verification stay sequential. Internal measurements showed 16% fewer round trips and 27% lower cost, and Bullet says that approach matters more than raw model speed. That’s the kind of claim that sounds boring until you’ve watched an agent burn half a day on avoidable chatter.
On SWE-bench Verified, Bullet says it solved 479 of 500 tasks, or 95.8%, in one attempt, averaging 119 seconds per task. It says that was 35–67% faster than mini-SWE-agent plus Fable or Sol, depending on the task. The team is also candid that code search itself was harder than expected, with regex-dialect mismatches causing silent misses and sending agents off course.
The broader use case, at least so far, seems to be long iterative work: benchmarks, data pipelines, and evaluation loops, where one step depends on the last and throwing more agents at the problem doesn’t magically help. Bullet is pitching speed, but the sharper idea is restraint. Fewer round trips. Less garbage in the context. Less time waiting for the machine to think out loud like it’s billing by the word.
My take — AI-written commentary, not fact-checked reporting
This is the right enemy to pick: not “AI coding” in the abstract, but the stupid waiting and bloated context that make these tools feel like interns with a caffeine problem. A lot of agent hype is just a nicer wrapper on expensive indecision. Bullet sounds less magical, more practical, and that’s usually where the real product is hiding.
Read more about this at: codewithbullet.com
