How to use Claude to build a faster coding agent
Bullet's founders used Claude Code while building a coding agent focused on routing, search, context hygiene and fewer round trips. Their post reports internal savings and benchmark results.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Improve [coding agent repository] by examining model routing, targeted code search and context handling. Batch independent investigation, bound tool output and add fallbacks for search misses. Keep dependent edits and verification sequential, then evaluate changes using [measurement procedure].
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- adi1
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The founders launched Bullet and report using it daily. Their stated measurements include 16% fewer round trips, 27% lower cost and 479 of 500 SWE-bench Verified tasks resolved in one attempt.
- Human involvement
- The founders describe daily use, internal measurements and benchmark evaluation, but do not detail code review procedures.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Use Claude Code to develop the Bullet coding agent.
- Implement model routing, targeted search and bounded context handling.
- Add search fallbacks to address silent regex-related misses.
Bullet's founders describe building a coding agent after repeatedly waiting for Claude Code and Codex during earlier projects. They explicitly say they used Claude Code to improve on their Claude Code experience, with Bullet beginning as a side project.
Their development focused on model routing, targeted code and context searches, and context hygiene. Reported measures included bounding tool output, removing stale screenshots and avoiding repeated file reads. They also organized turns around batching independent investigation, making a focused edit and then performing focused verification.
Search behavior became a practical obstacle: regex dialect differences could silently miss relevant code and send agents toward incorrect conclusions. The founders responded with targeted search, fallbacks and bounded context. They describe reducing round trips as more important than model speed, with independent operations running concurrently and dependent edits and verification remaining sequential.
The post reports internal measurements of 16% fewer round trips and 27% lower cost. It also claims 479 of 500 SWE-bench Verified tasks resolved in one attempt, averaging 119 seconds. These are the founders' reported results, not independently verified measurements here.
Who it's for: Developers investigating orchestration and context handling in coding agents.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · adi1
Frequently asked questions
What changes did the founders emphasize?
They describe model routing, targeted searches, bounded tool output, removal of stale screenshots, fewer repeated reads and batching independent investigation.
What performance results does the source claim?
The founders report 16% fewer round trips, 27% lower cost and 479 of 500 SWE-bench Verified tasks resolved in one attempt at an average of 119 seconds. These are source claims.
More details
Remaining steps
- Measure internal round trips and cost, then run SWE-bench Verified.
About this account
The founders used Claude Code during agent development; the source does not describe how much implementation proceeded without intervention.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=49283063. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: adi1 on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · Bullet · SWE-bench Verified
Coding agents · Context management · Code search
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.