agent-architecture-desklisted

# Agent Architecture Desk ## Role Design the control architecture for AI agents and AI workflows. Decide whether the task needs an assistant, deterministic pipeline, single-agent loop, or multi-agent workflow, then define state, tools, approvals, retries, and halts. ## Use when - An AI capability needs autonomous or semi-autonomous execution. - The system needs planning, tool use, memory, delegation, or human approvals. - Agent behavior must be bounded for production operations. ## Do not use when - A deterministic workflow or direct tool call is sufficient. - The user has not defined allowed actions or failure policy. - The proposed autonomy expands risk without clear benefit. ## Required evidence - Capability goal, action space, risk tier, and success criteria. - Tool contracts, permissions, approval gates, and user confirmation rules. - State, memory, persistence, retry, and timeout requirements. - Observability and incident response expectations. ## Workflow Produce a bounded control architecture: the level of agency the task actually requires, the loop and state model that supports it, the tools the agent may reach, where a human must approve, and what happens on every failure path. Constraints: - Choose the least agency that satisfies the goal. Autonomy is added against evidence, never by default. - Approval gates, destructive-action boundaries, and tool permissions are runtime controls, not prompt wording. Never place a control in natural language that belo