Dynamic Workflow

Opening

A dynamic workflow is an agent-orchestration primitive, popularized by Anthropic's Claude Code in 2026, that lets a single model instance fan its work across dozens, hundreds, or thousands of sub-agents running in parallel inside a sandboxed runtime. Boris Cherny described dynamic workflows at Y Combinator as "essentially an algebra for agents" — a small set of combinators (run agents in sequence, run agents in parallel, fan out and aggregate) that compose into complex multi-stage pipelines the model orchestrates on its own rather than being pre-scripted by humans.

Mechanism

Claude Code's dynamic workflows run inside the Bun JavaScript runtime, which Anthropic uses as a sandbox to spawn isolated virtual machines per workflow. The model decides the structure dynamically: it might run a first-pass batch of agents to explore the problem space, a second-stage batch to verify the first stage's outputs, and a third-stage aggregation. Cherny's example is rewriting the Bun runtime from Zig to Rust — one prompt, dynamic workflow, 11 days of orchestration. The Bun/Zig-to-Rust project is now in production: Claude Code runs on the rewritten runtime. Cherny frames dynamic workflows as a new form of test-time compute — distinct from training-time compute and from naive parallel sampling, because the model itself decides the parallelism structure rather than having it imposed externally.

Tradeoffs

Dynamic workflows require a sandbox that can cheaply spawn many agents (Bun + VM is Anthropic's choice; the cost structure differs across providers). The architecture pushes token spend up dramatically — Cherny describes Claude Code now running "hundreds of agents every day, sometimes thousands" maintaining its own codebase. The trade-off is that work that previously required dozens of engineers (dead-code cleanup, abstraction deduplication, test coverage repair) becomes a daily cron-driven routine the model handles autonomously. The failure mode is unbounded execution — workflows that should have terminated after a stage but keep spawning agents; Cherny's example of his own Electron-to-Swift rewrite running for 14+ days illustrates the trade-off between letting the model work as long as it wants and ensuring termination guarantees.

Open Questions

Whether dynamic workflows generalize beyond coding (Cherny suggests "really in-depth data analysis over some really complicated data" and "build a very complex feature that takes multiple stages" as candidates). How to verify outputs from a multi-stage orchestration without re-running the whole tree. Whether dynamic workflows compose recursively — a workflow that itself spawns workflows — or whether this produces coordination pathologies. Whether the "algebra for agents" framing becomes a programming-language-level abstraction (so workflows are written by humans and orchestrated by the model) or stays as the model improvising the structure at runtime.