Bullet just hit 95.

Waiting for a coding agent to finish a task is the ultimate productivity killer. Most of us have been there—watching the cursor blink while the agent loops through the same three files, wasting tokens and time. Bullet (a YC S26 company) basically exists because the founders got tired of the sluggishness of tools like Claude Code and Codex. Instead of just accepting the lag, they built a faster alternative that actually delivers.

What's impressive isn't just the speed, but how they're handling the LLM agent workflow to avoid the typical pitfalls of "agentic" coding. They aren't just wrapping an API; they've implemented some specific architectural choices to keep things lean.

How Bullet optimizes the AI workflow

The core philosophy seems to be "aggressive hygiene" and smarter routing. Here is the breakdown of their approach:

• Smart Model Routing: Instead of blindly hitting the most expensive model, it routes tasks based on complexity. It prevents the frustration of using a high-end model for a task that a smaller, faster one like Sonnet could have handled in a fraction of the time.
• Targeted Context Search: They've ditched the "embed the whole repo" approach, which often leads to noise. Instead, they use optimized greps for code and context search, ensuring the model only sees what it actually needs.
• Context Hygiene: This is huge. They bound tool outputs and purge stale screenshots or redundant file reads. By keeping the context window clean, the model doesn't get bogged down by its own previous "garbage" output.
• Efficient Turn Management: Rather than a chatty back-and-forth, Bullet batches independent investigations and performs surgical edits followed by a single focused verification. This reduced round trips by 16% and cut costs by 27%.

The results on SWE-bench Verified are wild. Bullet resolved 479 out of 500 tasks in one attempt, averaging 119 seconds per task. That's roughly 35% to 67% faster than the mini-SWE-agent + Fable/Sol combo depending on the specific task.

For anyone looking for a practical tutorial on how to integrate this into their stack, the team has focused heavily on the "Flash" aspect of the experience—essentially removing the friction between the prompt and the PR. If you're tired of your current AI workflow feeling like a slow crawl, this is a deployment worth checking out.

If you want to see the actual benchmarks and the methodology they used to hit those numbers, you can find the details here:

https://www.codewithbullet.com/blog/benchmark-results.html

It's refreshing to see a tool that prioritizes latency and context precision over just adding more "agentic" layers that usually just slow things down.