agent-loopslisted

# Agent Loops Design the smallest durable system that can learn, resume, stop, and prove what happened. This skill owns loop architecture; [product-loop](../product-loop/SKILL.md) owns product discovery and launch decisions. ## Workflow 1. **Define the decisions**: State the objective, evidence that can change direction, authority, budget or horizon, external side effects, and hard stop conditions. Separate scientific or task success from operational completion. 1. **Make the goal the driver**: Use the active harness's native `/goal` when the user requests a sustained goal. Define whether completion means an evidenced objective, a finished time window, or an objective bounded by a deadline. Follow [goals, prompts and continuation](references/goals-and-prompts.md); verify actual host capabilities instead of implementing another goal engine. 1. **Challenge the topology**: Keep a layer only when it closes a distinct feedback horizon. Merge any layer that merely renames another role. Read [loop topology](references/topology.md) when several decision horizons are needed. 1. **Choose minimal durable state**: Prefer a human-readable portfolio checkpoint, one checkpoint per work item, an append-only event log, concise sourced lessons, and derived reports. Give every fact one owner; do not create a second task database. For explicitly assigned concurrent sessions, record disjoint work scope, session identity, shared writer and account-wide resource allocation before writes. 1. **Wr