stas2-git Agent Architecture Methods
Use this skill to choose practical agent architecture patterns and convert intent into a buildable specification.
Workflow
- Clarify the target outcome, user workflow, risk level, and available tools.
- Choose the smallest architecture that can satisfy the goal:
- single agent with tools
- agent with skills
- orchestrated workflow
- multi-agent handoff
- event-driven or human-in-the-loop system
- Read references based on the design question:
- Broad architecture patterns:
references/agent_building_methods_extraction.txt - Intent-driven development and harness engineering:
references/day1_vibe_coding_intro.txt - Tools, MCP, and interoperability:
references/day2_agent_tools_interop.txt - Spec-driven production development:
references/day5_spec_driven_production_dev.txt
- Broad architecture patterns:
- Produce a system design with boundaries, contracts, state, tools, evaluation, security, and deployment assumptions.
- Prefer a concrete spec over abstract advice.
Design Checklist
- Inputs, outputs, and user-visible success criteria
- Tool and data authority boundaries
- Deterministic code versus model judgment
- State and memory model
- Human approval points
- Observability and evaluation cases
- Failure modes and rollback path
Evaluation Prompts
- Positive: "Design an agent that reviews expense events and asks a manager for approval above a threshold." Expected: architecture with events, policy, LLM role, HITL, state, and evals.
- Positive edge: "I have a vague idea for a research agent with web tools; make it production-ready." Expected: clarify boundaries, tool authority, safety, and evals before implementation.
- Negative: "Write the Python function for this already-designed helper." Expected: no architecture skill unless design choices are still open.