EN
Back to the archive

The archive · AI & Models · Product decision · 2026

Bullet bets coding-agent speed, not model capability, wins developers

Two Yale seniors, ex-AppLovin and ex-Citadel, launch a YC S26 coding agent that routes, parallelizes, and searches leaner to beat Claude Code on latency.

Bullet

The betThat the model is not the bottleneck: a tighter harness - routing, parallel tool calls, targeted search - makes agents faster, and speed is the moat.Live

What the business is

A speed-focused CLI and desktop coding agent that routes straightforward tasks to fast models, runs independent tool calls in parallel, and searches code without embedding the whole repository.

How it started

Adi and Alex, computer science seniors at Yale, founded Bullet after stints at AppLovin and Citadel. Their Launch HN describes starting with an AI hedge fund and pivoting five more times before noticing the same final boss: waiting hours for Claude Code and Codex to grind through their codebases, so they built a leaner agent loop themselves.

What happened

Bullet joined Y Combinator's Summer 2026 batch. On SWE-bench Verified it resolved 479/500 tasks (95.8%) in one attempt, averaging 119 seconds per task and 35-67% faster than a mini-SWE-agent baseline with Fable/Sol, with 16% fewer round trips and 27% lower cost. It launched a free private beta with a CLI installable from npm for macOS and Linux.

How it ended up

Still live in private beta: free to use with no subscriptions, positioned as a faster alternative to Claude Code and Codex that can plug into existing plans or raw API keys.

Background

Bullet is a Y Combinator Summer 2026 startup building a speed-first coding agent. Its founders, two Yale computer science seniors who previously worked at AppLovin and Citadel, built it after months of waiting for Claude Code and Codex to finish runs across six of their codebases. Their thesis: the models are fast enough, and the agent loop around them is the bottleneck.

The product makes three architectural bets: route straightforward work to fast models and escalate only when a task demands it, run independent tool calls in parallel instead of sequentially, and use targeted code search rather than embedding the entire repository into context.

On SWE-bench Verified, Bullet reported 479/500 tasks resolved (95.8%) in one attempt, averaging 119 seconds per task and claiming 35-67% speed gains over a mini-SWE-agent baseline, with 16% fewer round trips and 27% lower cost.

Its August 13, 2026 Launch HN reached 121 points and 89 comments on the Hacker News front page, and the same week it reached #5 on Product Hunt with 236 upvotes. The product launched as a free private beta CLI for macOS and Linux, able to use a user's existing Claude Code or Codex plan or raw API keys.

What has to be true

  • Coding agents spent two years competing on capability, so latency became the open wedge: simple tasks still take 30 seconds of file reading and context prep.
  • Because every agent wraps the same frontier models, the harness - routing, parallel tool calls, search strategy - is where the remaining seconds live.
  • Publishing a strong benchmark (95.8% on SWE-bench Verified) gives a two-week-old product instant credibility against established tools.
  • Free private beta with bring-your-own-key removes the switching cost and lets developers verify the speed claim without a new subscription.
  • The six-pivot backstory shows the founders kept the team and the frustration but dropped the ideas until they found the one they lived daily.

What can be applied

When every competitor uses the same models, the harness around them is the product: Bullet competed on latency, published benchmarks, and made architecture - not a new model - the differentiator.

Aftermath

As of September 2026, Bullet remains in free private beta as a CLI and desktop app for macOS and Linux, installable via npm, and is positioned as the fast alternative to Claude Code and Codex. Independent reviewers note the benchmark comparison is not perfectly symmetric and the product is too new to verify at scale, but the architectural approach of routing, parallel execution, and targeted search is regarded as sound regardless of the exact score.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases