-
Notifications
You must be signed in to change notification settings - Fork 0
Open
Description
Summary
When Groundwork orchestrates multi-agent workflows (exploration, design, planning, implementation), there's no skill governing which model to dispatch for each task type. Currently, all subagents inherit the parent model by default, which wastes tokens on tasks that don't need frontier-level reasoning.
Problem
- Exploration tasks (file search, pattern discovery, codebase survey) may not need Opus — Haiku or Sonnet could be sufficient and dramatically cheaper.
- Design and synthesis tasks (document writing, architectural decisions, plan creation) benefit from Opus.
- Without explicit routing guidance, every subagent runs at the same cost regardless of task complexity.
- Token budget is finite. Right-sizing model selection per task type stretches the budget toward higher levels of co-creative work.
Desired Outcome
A skill that provides model-routing guidance for subagent dispatch. It should define:
- Task-type to model-tier mapping — which kinds of work warrant which model class
- Decision criteria — how to assess whether a task needs frontier reasoning or can be handled by a faster/cheaper model
- Override rules — when to escalate (e.g., exploration that turns out to need synthesis)
Open Questions
- What's the right model tier for exploration tasks? (Haiku? Sonnet? Needs experimentation)
- Should this be a standalone skill or integrated into
using-groundworkas a routing extension? - How does this interact with
subagent-driven-development, which already dispatches subagents?
Acceptance Criteria
- Skill defines task-type categories with recommended model tiers
- Routing guidance is actionable (agent reading it knows which model to request)
- Covers at least: exploration, design/synthesis, implementation, review, debugging
- Documents when to override the default mapping
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels