loading...
Bullet parallelizes independent tool calls – searches, reads, verifications – to run 30-60% faster than Claude Code and Codex. Free. No subscription of its own.
ENTRY ANGLES
Vertically specialized coding agent for high-frequency CI/CD workflows · Coding agent performance benchmarking tool for enterprise workflow samples · Parallel execution layer as open-source infrastructure for any coding agent
VERTICALS
CAPABILITIES
LLM orchestration and routing, Tool call parallelization, Code search and indexing, Performance benchmarking and reporting
Coding agents have converged on a fundamental architecture. They receive a task, search the codebase, read relevant files, plan a solution, write the code, run tests, and iterate. The problem is that a significant fraction of those operations are logically independent of one another: a search across three directories and a read of two configuration files can all happen simultaneously without affecting each other's results. Essentially every major coding agent runs them sequentially anyway. That is not a constraint of the problem; it is a design choice that nobody has revisited until now.
Bullet, a YC S26 company built by Yale CS graduates who previously worked at AppLovin and Citadel, is the revisit. The agent parallelizes independent tool calls – searches, reads, verifications that don't depend on each other's results – rather than queuing them. It intercepts duplicate calls and stuck loops before they waste time. It auto-selects the reasoning depth appropriate for each prompt rather than defaulting to maximum reasoning on every query. The result, the company claims, is 30-60% faster task completion versus Claude Code and Codex.
SWE-bench Verified offers partial confirmation: Bullet scores 95.8%, placing in the top three on that leaderboard at 119 seconds per task. The product is free, requires no separate subscription, runs on top of existing Claude Code or Codex subscriptions or API keys, and is available as an npm package for macOS and Linux.
The speed claim survives scrutiny at the architectural level. The core insight – that independent operations in a coding agent's workflow can be parallelized – is not Bullet's proprietary discovery. It is an accurate description of a structural inefficiency that competing agents have deprioritized in favor of simplicity and determinism. Sequential execution is easier to debug, easier to make reproducible, and easier to explain when something goes wrong. Bullet is betting that developers who watch Claude Code spend four minutes completing a task that takes Bullet 90 seconds will tolerate somewhat less debuggability in exchange for the time.
That trade is not equally valid for all tasks. For long searches across large codebases – the cases where sequential execution is most wasteful – the performance gap is plausible and meaningful. For tasks where correctness and reasoning quality matter more than wall-clock time, the advantage narrows to the vanishing point. Bullet is a tool for speed-sensitive workflows, not a universal upgrade.
The free-tier, open distribution strategy maps to a specific playbook: land developers through speed, build usage patterns and stickiness, convert through a premium tier. Claude Code used an aggressive free tier to displace GitHub Copilot from a significant share of professional developer workflows in 2025. Bullet is running the same playbook at the next layer of abstraction, competing not against the underlying AI providers but against the specific UX decisions their agents have made.
Bullet's model confirms a pattern: the productivity gains from parallelizing agent operations are large enough to sustain an independent product even when the underlying AI capability comes from third-party models. The same architecture could be applied to any multi-step AI workflow where independent operations are currently executed sequentially – data analysis pipelines, content production workflows, research synthesis agents. The architectural insight is general; Bullet is applying it only to coding.
The more defensible build, for someone entering the coding agent space now, is not a faster generalist agent but a faster agent specialized for a single class of problem. Bullet competes across the full range of development tasks. A competitor that is 3x faster specifically for refactoring tests across large codebases – because it has been purpose-built to parallelize that specific operation rather than every operation simultaneously – has a clearer performance claim and a more defensible niche.
Enterprise CI/CD pipelines are where the architectural advantage is most commercially significant. Organizations running dozens of automated coding tasks per day have dramatically higher speed sensitivity than individual developers doing occasional feature work. The gap that creates a product opportunity: there is no established benchmarking tool that measures coding agent performance against an organization's actual workflow sample rather than academic benchmarks. A builder who produces that tool creates a distribution channel for any agent that outperforms the baseline – and acquires the data needed to identify which category of task is most worth specializing for.