Why Ryan Dahl Built Deno and Celld: An AI‑Agent Swarm Investigation

Curious why Node.js creator Ryan Dahl later launched Deno and the newer Celld, the author tasked EvoX’s Cowork "swarm" mode with a comprehensive investigation of the three runtimes.

First round: five agents probe Node.js, Deno and Celld

EvoX broke the work into six steps—separate sub‑tasks for each project, an independent cross‑check of the claim "Celld is the third‑generation Node.js", and final HTML and CSV deliverables—so that conclusions in one branch could not bias the others.

The system generated a plan, dispatched five sub‑agents, and collected raw evidence, sources, and divergent opinions.

Task‑organization modes

Single mode : one agent handles the whole problem in a single context.

Sub‑Agent mode : a master agent splits the job, then reads, compresses, and rewrites each sub‑agent’s output.

Swarm mode : each sub‑task runs in an isolated context with a stable identifier; results are merged without extra reinterpretation.

Swarm mode dramatically cuts long‑context costs. In a benchmark of 256 equally sized sub‑tasks, a single long context consumed about 49.34 million tokens, while the swarm approach used only 1.04 million tokens—a roughly 48‑fold reduction. The first round ran for 10 minutes 11 seconds, producing useful analysis frameworks but leaving some primary sources for Celld incomplete.

After supplementing missing material (celld.dev, GitHub README, Ryan Dahl’s talks), EvoX delivered two artifacts: a 38 KB HTML report containing timelines and comparison tables, and a 39‑row CSV that traced every conclusion back to its source.

The report placed each project in its era: Node.js tackled web‑backend concurrency; Deno responded to Node’s security and engineering shortcomings; Celld targets self‑hosted, stateful Durable Objects. Crucially, the author disproved the hypothesis that "Celld is the third‑generation Node.js"—no official evidence supports that classification, and the three projects solve distinct problems despite shared lineage.

Memory reuse and self‑evolution

EvoX reused four prior memory entries, triggering the builtin:demand_insight capsule with a 66 % local‑memory reuse rate, illustrating the platform’s ability to treat past work as reusable experience rather than mere chat history.

Second round: adding Bun without restarting

To test reuse, the author asked EvoX to extend the investigation to Bun while preserving earlier results. EvoX invoked recall_assets, fetched six relevant assets from memory/evolution/swarm, and launched three new Bun sub‑tasks.

The system updated the original HTML in place—adding a Bun navigation section, timeline, four‑column comparison, cross‑check, and a “what was reused” block—growing the file to 56 KB (nine chapters, nine tables) and expanding the CSV to 49 rows.

The updated timeline marks Bun as a 2021‑2022 side‑branch, clarifying that Bun shares no author, organization, engine, or language lineage with Ryan Dahl’s projects; it simply pursues Node‑compatible, full‑stack performance.

Takeaways and suitable tasks

The two rounds illustrate two complementary strengths: Swarm mode excels at organizing complex, long‑context tasks by partitioning work, while the self‑evolution layer reuses verified results for subsequent investigations. Ideal use cases include multi‑project technical surveys, evidence‑backed research requiring cross‑validation, large‑scale document analysis, and any long‑running task that must output web pages, tables, or reports.

EvoX remains in beta, free to use without invitation. Users with opaque, deliverable‑heavy problems are encouraged to try it and contribute feedback for further development.