Agent Arch

Agent loops, memory, tools, and safety boundaries. Use when designing AI agents.

Agent Architecture Skill

This skill provides structured guidance for Agent Architecture work. Act as an active guide: confirm triggers, propose the stages below, and adapt if the user wants a lighter pass.

When to Offer This Workflow

Trigger conditions:

• User mentions agent architecture or closely related work
• They want a structured workflow rather than ad-hoc tips
• They are preparing a review, rollout, or stakeholder communication

Initial offer: Explain the four stages briefly and ask whether to follow this workflow or work freeform. If they decline, continue in their preferred style.

Workflow Stages

Stage 1: Clarify context & goals

Anchor on goals, tools, and constraints. Ask what success looks like, constraints, and what must not break. Capture unknowns early.

Stage 2: Design or plan the approach

Translate goals into a concrete plan around memory and state. Compare alternatives and explicit trade-offs; avoid implicit assumptions.

Stage 3: Implement, validate, and harden

Execute with verification loops tied to planning loops and stopping. Prefer small steps, measurable checks, and rollback points where risk is high.

Stage 4: Operate, communicate, and iterate

Close the loop with safety monitoring: monitoring, documentation, stakeholder updates, and lessons learned for the next cycle.

Checklist Before Completion

• Goals and constraints are explicit for Agent Architecture Skill
• Risks and trade-offs are stated, not hand-waved
• Verification steps match the change’s impact (tests, canary, peer review)
• Operational follow-through is covered (monitoring, docs, owners)

Tips for Effective Guidance

• Be procedural: stage-by-stage, with clear exit criteria
• Ask for missing context (environment, scale, deadlines) before prescribing
• Prefer checklists and concrete examples over generic platitudes
• If the user declines the workflow, switch to freeform help without lecturing

Handling Deviations

• If the user wants to skip a stage: confirm and continue with what they need.
• If context is missing: ask targeted questions before strong recommendations.
• Prefer concrete examples, trade-offs, and verification steps over generic advice.

Quality Bar

• Each recommendation should be actionable (what to do next).
• Call out failure modes relevant to Agent Architecture (security, scale, UX, or ops).
• Keep tone direct and respectful of the user’s time.