Show HN: Shard – Stop watching one AI agent code for 45 min. Run four at once
TL;DR
Shard automatically decomposes a large coding prompt into a DAG of parallel sub-tasks.
Key Points
- Each sub-task receives exclusive file ownership, eliminating merge conflicts by design.
- Multiple agents run simultaneously in separate git worktrees and are merged in topological order.
- Test failures are self-healed automatically. A 45-minute serial task shrinks to ~12 minutes with 4 parallel agents.
- Open source, cross-platform, works with Claude Code, Aider, and Cursor.
Nauti's Take
The approach is smarter than it looks at first glance: not just 'launch many agents', but model the dependencies first, then execute in parallel. That is the difference between a distributed system and a chaotic pile of threads.
The open question is how well the automatic task decomposition handles truly complex, tightly coupled codebases – a DAG is only as good as its edges. Still, anyone working daily with long agent runs should keep a close eye on this project.
Context
The real bottleneck in AI-assisted coding is not model quality but serial wait time – you start a task and are blocked for an hour. Shard addresses this with a structured parallelization approach rather than naive multi-agent chaos. The DAG model with exclusive file ownership solves the core problem of coordinated parallel code changes in a technically clean way.
If the promise holds, it shifts the productivity curve for developers significantly – not through better models, but through better orchestration.
Sources