Dynamic workflows, as pioneered by @omarsar0, generate task harnesses on the fly for agent orchestrators. The approach enables branching, parallel, and verified agent tasks across coding agents like Claude Code and Codex.
Key facts
- Dynamic workflows generate task harnesses on the fly.
- Applied to 10+ use cases including branching research and bug hunting.
- Works with Claude Code, Codex, Pi, and custom agents.
- Monitoring dashboard built as an HTML artifact.
- Extends beyond coding to business and science domains.
Dynamic workflows are emerging as a new primitive for agent orchestration, enabling on-the-fly generation of task harnesses. According to @omarsar0, the concept feels as foundational as agent skills, incorporating dynamic behaviors, cooperation, and verification into complex, long-running tasks.
The approach has been successfully applied to a range of use cases: branching deep research tasks with verification, parallel deep research tasks, session mining of all agent sessions, bug hunting, triaging, fact-checking, LLM councils, AI simulations, data synthesis, and evaluations generation. @omarsar0 built a monitoring dashboard as an HTML artifact to track tasks, metrics, and reports for his custom agent orchestrator.
Crucially, the concept is not limited to coding tasks. @omarsar0 notes it extends to business use cases and technical domains like science and research. The ability to generate harnesses on the fly and integrate monitoring directly into the workflow represents a shift from static, predefined agent pipelines to dynamic, adaptive orchestrations.
The unique take: Dynamic workflows address a structural weakness in current agent architectures — the inability to adapt task decomposition at runtime based on intermediate results. By generating harnesses dynamically, the orchestrator can branch, verify, and parallelize without pre-scripting every path. This is a meaningful departure from the rigid DAG-based approaches common in tools like LangGraph or Prefect.
As @omarsar0 puts it: "There is so much exploration ground here." The key question is whether dynamic workflows will be adopted as a standard primitive by major agent frameworks or remain a bespoke pattern for advanced users.
What to watch
Watch for whether major agent frameworks (LangChain, Vercel AI SDK, CrewAI) adopt dynamic workflows as a first-class primitive in their next releases, or if @omarsar0 open-sources his orchestrator implementation.
