Claude Dynamic Workflows

Dynamic Workflows is a research-preview feature of Claude Code itself — Anthropic's own multi-agent orchestration mechanism, not something MadoHub built. Claude Code can fan a single session out into many concurrent subagents working on the same problem.

MadoHub ships a compatibility layer on top of it. MadoHub does not create or run these subagents — it detects them, keeps its own agent-state model correct while a workflow is running, and gives you a way to see progress without the flood of noise a naive integration would produce. Without that layer, a workflow with 65 subagents would mean 65 completion badges stacking up on one tile, each needing to be dismissed individually.

You reach Dynamic Workflows from inside a Claude Code session by typing ultracode in a prompt, running /effort ultracode, or flipping the Workflows toggle in /config on a Pro plan. It requires a recent Claude Code version; the natural-language trigger needs an even newer one.

What MadoHub does with it

• Detects when a workflow is running on a Claude Code tile and aggregates subagent events into a single per-tile progress counter instead of one badge per subagent.
• Shows progress in three places — the Context Gauge's active indicator (a small pulsing dot) lights up on the tile, the Fleet panel shows a workflow · 3/5 agents line under the row, and the Context Gauge tooltip shows an aggregated workflow token figure.
• Keeps workflow token usage separate from the main gauge percentage — subagents run in their own contexts, not the main session's window, so their token usage is tracked apart from the main session's fill.
• Resets counters when a new run starts on the same tile, so a second workflow does not inherit the first one's tally.

How to use it

You do not "use" MadoHub's compatibility layer directly — you use Claude Code's Dynamic Workflows feature inside a Claude Code tile, and MadoHub surfaces the progress automatically.

1. Open a Claude Code tile and trigger a workflow from inside the session (typing ultracode, running /effort ultracode, or flipping the Workflows toggle in /config).
2. Watch the tile's Context Gauge active indicator light up while the workflow runs.
3. Check the Fleet panel for a workflow · {done}/{seen} agents line under that tile's row to see fan-out progress without flying to the tile.
4. Hover the tile's Context Gauge for the aggregated workflow token figure, appended separately from the main percentage.
5. When the workflow finishes, the active indicator turns off and the progress line disappears automatically.

Claude's own limits

These are Claude Code's limits, not MadoHub's, and MadoHub cannot change them:

Common use cases

• "Run a big research task across many subagents." — trigger a workflow inside Claude Code and watch fan-out progress from the Fleet panel without leaving your work.
• "Is that workflow still active?" — the Context Gauge's pulsing dot and the (active) suffix in the tooltip tell you at a glance.
• "How much token burn is the workflow generating?" — hover the Context Gauge; the workflow token figure is shown separately from the main session's fill.

Tips and best practices

• Watch the Fleet panel's workflow · {done}/{seen} agents line for fan-out progress instead of flying to the tile — the whole point of the compatibility layer is to keep you from drowning in per-subagent badges.
• The workflow token figure is tracked separately from the main gauge percentage on purpose. Do not add them together to estimate "total context" — subagents run in their own contexts.
• A freshly launched workflow can show as active with zero tokens tallied yet, before its first usage block lands. Give it a moment.
• Once a workflow finishes, Claude Code writes an authoritative total; MadoHub reads that instead of re-summing subagent files, so the final figure is more accurate than the running one.

Related features

• Context Gauge — the per-tile fill bar whose active indicator and tooltip surface workflow progress.
• Fleet — the panel that shows workflow · {done}/{seen} agents per row.
• MadoAgent — for MadoHub's own multi-agent patterns (Ensemble Exploration, Orchestrator Mode).

Limitations

• Dynamic Workflows is Claude Code's research-preview feature. MadoHub detects and visualizes it; MadoHub does not implement it.
• Everything MadoHub reads here is observed behavior Claude Code writes for its own purposes, not a published contract. Anthropic can change the format without notice as the feature moves out of research preview.
• User input is not accepted while a workflow is running — only a permission prompt can pause a run. Plan accordingly before triggering one.