docs(rfc): feedback on dynamic workflows RFC

[dcieslak19973]

This adds a short review of the dynamic workflows RFC (added in #3431) as a companion doc in rfc/.

The v1 scope looks right: a bounded, single-player, deterministic agent-orchestration engine. This feedback does not argue for adding features to v1. It proposes keeping four abstractions one notch more general than v1 strictly needs, so later use cases can ride the same engine without a rewrite:

1. External-input steps — let a durable step resolve to a recorded result supplied by a human or an external event, not only an agent task. Unlocks human-in-the-loop approval, after-action reviews, and bounded incident investigation later.
2. Addressable run identity + event bus — keep runs observable independent of the launching chat (the RFC already anticipates dashboard/list and background views).
3. Determinism boundary, written down — affirm deterministic replay, and note that journaling external inputs is what keeps reactive-but-bounded runs inside it.
4. Pluggable AgentExecutor — keep the execution path behind an interface (mux TaskService as the sole v1 implementation) so non-mux backends (e.g. Anthropic Managed Agents) can fulfill steps later. A clean decoupling/testability win regardless.

Each ask is near-zero v1 build cost, and the doc explicitly lists what it is not asking for.

Only adds rfc/20260529_dynamic-workflows.feedback.md — no code changes.

🤖 Generated with Claude Code