Technology Radar
Where a team of coding agents is a small, deliberate group, a coding agent swarm applies dozens to hundreds of agents to a problem, with AI determining composition and size dynamically. Projects such as Gas Town and Ruflo (formerly Claude Flow) are good examples of this approach. Early patterns for swarm implementations are emerging: hierarchical role separation (orchestrators, supervisors and ephemeral workers), a durable work ledger that helps agents divide and coordinate work (Gas Town uses beads for this) and a merging mechanism to handle conflicts from parallel work.
Two swarm experiments have drawn particular attention: Anthropic's C compiler generation and Cursor's agent scaling experiment that created a browser over a week. It's worth noting that both teams chose use cases that could rely on existing detailed specifications, and in the case of the C compiler, comprehensive test suites that provide clear, measurable feedback. Those conditions are not representative of typical product development, where requirements are less defined and verification is harder. Nevertheless, these experiments contribute to emerging patterns for making long-running swarms technically viable. They remain costly and are still far from mature, which is why we advise caution when adopting this technique.
