Evo, the open-source autoresearch plugin for Claude Code, shipped a port of its six-step research orchestrator onto Anthropic’s dynamic workflows, moving phase logic, fan-out width, stopping rules, and gate checks from in-context prompt text into scripted JavaScript that subagents execute with fresh scoped context.
The practical change resolves long-horizon instruction adherence. Previously, carrying the full research method through every subagent invocation ballooned context usage and let orchestrator drift corrupt later phases. Now the model makes judgment calls (which hypothesis to pursue, when to stop) and the code handles coordination (parallel fan-out, regression locking, gate execution). The method is the code.
Anthropic introduced the dynamic workflow primitives in Claude Code on June 2, enabling JavaScript orchestration scripts that spawn up to 1,000 parallel subagents per run. Evo’s update, published June 9 on GitHub at github.com/evo-hq/evo, is an early open-source reference implementation of the same split applied to research workflows.
Teams building production agent pipelines that need to run past a single context window should study this architecture before committing to prompt-only orchestration.
Evo (github.com/evo-hq/evo), published 2026-06-09.
