What problem does this solve?
Complex tasks currently run in a single agent context. This makes long workflows harder to structure, reduces specialization, and increases token waste when one prompt has to do planning + coding + verification together.
Proposed solution
Add a subagent orchestration layer similar to Claude Code style workers:
- Parent agent delegates scoped tasks to specialist workers (Planner, Coder, Reviewer).
- Each worker runs with isolated short-lived context and returns a concise report.
- Parent agent merges outputs and decides next step.
- Add guardrails: max worker depth, timeout, and per-worker tool allowlist.
Suggested implementation surfaces:
- �gent/subagents/ package for orchestration logic
- Internal API like
un_subagent(task, role, tools, context)
- Telemetry events in session logs for worker start/end
Alternatives considered
Keep a single-agent loop and rely only on prompt engineering. This is simpler, but gives weaker separation of concerns and poorer debuggability for multi-step tasks.
What problem does this solve?
Complex tasks currently run in a single agent context. This makes long workflows harder to structure, reduces specialization, and increases token waste when one prompt has to do planning + coding + verification together.
Proposed solution
Add a subagent orchestration layer similar to Claude Code style workers:
Suggested implementation surfaces:
un_subagent(task, role, tools, context)
Alternatives considered
Keep a single-agent loop and rely only on prompt engineering. This is simpler, but gives weaker separation of concerns and poorer debuggability for multi-step tasks.