A new Y Combinator-backed coding agent launched on Hacker News this week, and it's making a bold claim: 95.8% on SWE-Bench Verified — the benchmark for AI coding agents — while being dramatically faster than competitors. Bullet, from YC's Summer 2026 batch, is a coding agent built around one principle: your ideas move fast, your agent should too.
The Problem: Agent Latency
The Bullet team's pitch starts with a pain point every AI-assisted developer knows: waiting for agent runs. Not because the models are incapable, but because the machinery around them is heavier than it needs to be.
If you've used Claude Code, Cursor, or similar tools, you've experienced this: you ask the agent to do something simple, and it spends 30 seconds reading files, searching the codebase, and preparing context before it even starts generating code. For complex tasks, this is acceptable. For simple tasks, it's infuriating.
Bullet's thesis is that most coding agent tasks are straightforward, and the overhead of the agent framework — not the model — is the bottleneck.
Three Key Optimizations
Bullet's approach centers on three architectural decisions:
1. Smart Model Routing
Not every task needs the most powerful (and slowest) model. Bullet routes straightforward work to fast models and escalates only when the task demands it. This is the "router" pattern — use a cheap, fast model to classify the task, then route it to the appropriate model.
This is similar to how OpenAI's o1 works (thinking time proportional to problem difficulty), but applied at the agent level rather than the model level. The result: simple tasks complete in seconds, complex tasks get the full power of frontier models.
2. Targeted Search
Most coding agents embed the entire repository into context before starting work. This is expensive (token costs) and slow (embedding generation). Bullet uses targeted search and file reads to find relevant code without embedding the whole repo.
This is a significant architectural choice. Embedding-based search (like what you'd get with a vector database) is comprehensive but slow. Targeted search (using file structure, imports, and grep-like techniques) is faster but might miss things. The tradeoff favors speed, which aligns with Bullet's thesis that most tasks are straightforward.
3. Parallel Tool Execution
Bullet runs independent tool calls in parallel. If the agent needs to read three files, it reads all three simultaneously rather than sequentially. It also intercepts duplicate calls and stuck loops before they waste time.
This is where the biggest speed gains come from. Sequential tool calls are the hidden tax of most agent frameworks — each file read waits for the previous one to complete. Parallel execution can cut agent run time by 50-70% for tasks that require multiple file reads.
The SWE-Bench Score
95.8% on SWE-Bench Verified is a significant claim. For context:
- SWE-Bench Verified is a curated subset of SWE-Bench where human annotators verified that the issues have clear solutions and the tests correctly evaluate those solutions.
- Claude Code (Anthropic's official agent) scores around 70-75%.
- OpenAI's Codex scores in a similar range.
- 95.8% would put Bullet at the very top of the leaderboard.
However, there's an important caveat: SWE-Bench scores can be gamed. Some high scores have been achieved by allowing more attempts per problem, using more powerful models, or having human-in-the-loop verification. The SWE-Bench leaderboard has had issues with reproducibility and fair comparison.
Bullet is in private beta, so independent verification isn't possible yet. But the architectural approach — parallel execution, smart routing, targeted search — is sound and should produce meaningful speed improvements regardless of the exact benchmark score.
What This Reveals About the AI Dev Tool Market
Bullet's launch is symptomatic of several trends in the AI development tool market:
Speed Is Becoming the Differentiator
When all coding agents use the same underlying models (Claude, GPT, etc.), the differentiation comes from the infrastructure around them. Bullet is betting that speed will be the primary differentiator — developers will choose the agent that gets things done fastest, not the one with the most features.
The "Tight Loop" Philosophy
Bullet's architecture is a "tighter loop" around the same model → tools → results pattern. This is a design philosophy that prioritizes minimalism: fewer abstractions, less overhead, more direct execution. It's the Unix philosophy applied to AI agents.
Open Source vs. Proprietary
Bullet is offering free access during private beta, with a CLI available via npm install -g @trybullet/cli. This suggests a freemium model — free for individual use, paid for teams and enterprises. This is the same model used by most coding agent tools, including Cursor and Claude Code.
Should You Try It?
Bullet is available for macOS and Linux (Node 18+), which makes it accessible to most developers. The free private beta means there's no cost to try it.
For developers frustrated with agent latency — and who isn't? — Bullet's approach is worth testing. The parallel execution alone could be a significant productivity boost for tasks that require reading multiple files.
However, as with any new tool, the "choose boring technology" principle applies. If you're already productive with Claude Code or Cursor, switching to Bullet during its private beta means spending an innovation token on an unproven tool. It might be worth watching Bullet's development and waiting for it to mature before making it your primary agent.
The Bigger Picture
Bullet's launch signals that the AI coding agent market is entering its next phase. The first phase was about capability — can AI agents solve real coding problems? The second phase, which we're now entering, is about efficiency — can they do it fast enough to feel native?
When coding agents are as fast as typing, the way developers work changes fundamentally. You stop thinking about "asking the agent" and start thinking about "working alongside the agent." The latency barrier is the last thing standing between us and that future.
Bullet is one of the first agents designed from the ground up to cross that barrier. Whether it succeeds remains to be seen, but the direction is right.
Bullet is in private beta at codewithbullet.com. Available for macOS and Linux via npm install -g @trybullet/cli.
Top comments (0)