Ensemble Exploration: Try N Approaches at Once

Some problems don't have an obvious first approach. LRU vs LFU vs ARC for a cache. Two competing refactor strategies. A recursive implementation against an iterative one. Until now, exploring those options meant spawning agents one at a time and comparing them manually. Today we're shipping Ensemble Exploration — a way to fork a single problem into several divergent agents, watch them work in parallel, and compare the results in one place.

What you can do now

• Fork one problem into several approaches — describe the problem and the angles worth trying, and MadoAgent spawns a separate agent for each, all running in parallel.
• Watch every fork from one panel — the new Exploration tab in the right drawer lists each fork with a live status dot and lets you open a side-by-side comparison view as soon as output starts arriving.
• Score forks with one click — a single Score forks button asks a lightweight critic model to rate each fork on correctness, completeness, clarity, and actionability, and recommend one, with a one-line rationale.
• Pick the winner yourself — there's no auto-selection. You read the output and click Pick winner on the fork you want to keep.
• Have your runs survive a restart — ensemble runs are remembered, so closing and reopening MadoHub doesn't lose them.

How to try it

1. Press Cmd+Shift+M to open MadoAgent.
2. Ask it to explore a few approaches to a problem you're unsure about — for example, "try three different cache eviction strategies for the session lookup cache."
3. Open the Exploration tab in the right drawer to watch the forks land.
4. When they're done, click Score forks for a quick comparison, then Pick winner on the fork you want to keep.

Why it matters / use case

When the right answer depends on tradeoffs you can't fully evaluate until you see working code, running one agent and hoping it picked well is a gamble. Ensemble Exploration turns that gamble into a side-by-side comparison: three or four forks explore different regions of the solution space, and you choose based on actual output rather than a hunch. It's most useful for genuinely uncertain decisions — eviction policies, indexing strategies, migration shapes — where the cost of a few extra agent runs is small compared to the cost of picking the wrong approach and rewriting later.

Related docs

• Ensemble Exploration
• MadoAgent
• The Right Drawer

What's next / limitations

Forks are capped at five per run so ensembles stay focused on genuinely uncertain decisions rather than becoming a default multiplier. The critic is advisory only — it highlights a recommended fork with a badge, but selecting a winner is always your call, because a quick scoring pass can't know whether "clarity" should outweigh "performance" for your specific situation.