Cosine Swarm ships parallel subagents as the default mode, not a gimmick
Cosine just turned Swarm from a research demo into the default mode of their coding agent. One prompt launches a coordinated hierarchy, an Orchestrator breaks the goal down, Task Owners manage workstreams, Workers execute tactical changes in parallel isolated contexts. Everything runs across CLI, Desktop, and Cloud on the same runtime. No more single thread choking on a complex migration.
The claim that matters is their own benchmark, 3x or greater throughput gains on production-grade tasks versus single-agent. This is not the usual multi-agent paper that looks great on toy problems and falls over in real codebases. Cosine runs on actual enterprise refactors, and the architecture choice is that each subagent gets its own isolated context, its own tools, its own trajectory, so when something goes sideways in a Worker it does not corrupt the Orchestrator plan. It branches, it stays reviewable, it does not turn into chaos.
The editorial question this run has kept coming back to is why a dozen coding-agent startups look identical from the outside. Cosine Swarm is the clearest answer so far, because they shipped the thing most teams only write blog posts about. Cursor is still chasing a great IDE, Claude Code and Codex are first-party single-agent loops, Factory is betting on desktop. Cosine is the one saying the actual bottleneck is serial execution and burning a year of engineering to fix it at the runtime layer. If their 3x number holds on customer code, the rest of the market has to follow.
The second-order bet is that agent throughput becomes the buying criterion. Right now people pick coding agents by demo quality or IDE feel. Cosine wants you to pick by tasks-finished-per-day, which is the boring enterprise metric that eventually wins. Whether that narrative sticks depends on whether the Orchestrator can actually decompose novel work without hallucinating dependencies, which is where multi-agent systems usually die.
Available now at https://cosine.sh and the architecture writeup at https://cosine.sh/blog/parallelising-software-development-multi-agent-productivity.
← Back to all articles
The claim that matters is their own benchmark, 3x or greater throughput gains on production-grade tasks versus single-agent. This is not the usual multi-agent paper that looks great on toy problems and falls over in real codebases. Cosine runs on actual enterprise refactors, and the architecture choice is that each subagent gets its own isolated context, its own tools, its own trajectory, so when something goes sideways in a Worker it does not corrupt the Orchestrator plan. It branches, it stays reviewable, it does not turn into chaos.
The editorial question this run has kept coming back to is why a dozen coding-agent startups look identical from the outside. Cosine Swarm is the clearest answer so far, because they shipped the thing most teams only write blog posts about. Cursor is still chasing a great IDE, Claude Code and Codex are first-party single-agent loops, Factory is betting on desktop. Cosine is the one saying the actual bottleneck is serial execution and burning a year of engineering to fix it at the runtime layer. If their 3x number holds on customer code, the rest of the market has to follow.
The second-order bet is that agent throughput becomes the buying criterion. Right now people pick coding agents by demo quality or IDE feel. Cosine wants you to pick by tasks-finished-per-day, which is the boring enterprise metric that eventually wins. Whether that narrative sticks depends on whether the Orchestrator can actually decompose novel work without hallucinating dependencies, which is where multi-agent systems usually die.
Available now at https://cosine.sh and the architecture writeup at https://cosine.sh/blog/parallelising-software-development-multi-agent-productivity.
Comments