Agent, skill, or MCP? Which to use and when to use them | AWS Clare Liguori

The podcast discusses the evolution and challenges of agent development, emphasizing a shift from complex, over-engineered systems to simpler, model-driven architectures. As AI models have become more capable in reasoning and tool use, the need for intricate orchestrations and multi-agent frameworks has diminished. This has led to a growing trend of minimizing engineering overhead, avoiding unnecessary customization, and trusting the model's inherent abilities. Tools like the Strands Agents SDK emerged from this philosophy, focusing on developer-centric design by reducing cognitive load and enabling rapid deployment through declarative interfaces.

Organizations are increasingly prioritizing system-wide efficiency over individual productivity, adopting reusable, standardized agent systems that support observability, evaluation, and composability. Internal successes at AWS and similar companies highlight the benefits of consolidating fragmented, domain-specific agents into fewer, more generalized agents with progressively disclosed skills. Challenges remain in aligning organizational incentives and overcoming resistance to decommissioning obsolete systems. The discussion also covers advancements in AI code review, where policy-driven automation improves quality control, and the need for better infrastructure - such as evolved version control systems - to support future multi-agent collaboration.

• What if you rebuilt your agent with only one model, one prompt, and one tool?

  • Move: Strip down your current agent implementation to use just a single LLM call with a clear system prompt, one high-leverage tool (e.g., database query or API client), and no orchestration logic. Deploy this minimal version as a parallel test endpoint.
  • Why Now?: Modern models like Sonnet 3.5+ handle complex reasoning and tool use natively - your custom loops and state management may be degrading performance. With new model releases every 6 months, now is the time to test if simplicity outperforms your existing stack.
  • Expected Upside: 3 - 10x faster iteration cycles, reduced debugging surface, and higher reliability by removing engineering overhead that's now obsolete.

• What if you published a reusable skill for your core product API instead of building another agent?

  • Move: Define a standardized skill descriptor (in MCP or similar format) for your most-used API workflow (e.g., "create customer invoice"), host it publicly, and integrate it into a lightweight CLI or Slack bot using a simple agent framework like Strands.
  • Why Now?: The industry is shifting from isolated agents to composable skills. Tools like MCP are being simplified for HTTP-based, stateless usage - making it easier than ever to publish and distribute skills without full agent infrastructure.
  • Expected Upside: Faster adoption across teams and customers, reduced duplication of integration logic, and positioning your API as a first-class citizen in agentic workflows.

• What if you replaced your next code review with an AI-first pre-check pipeline?

  • Move: Build a pre-submission hook (e.g., Git commit-msg or PR action) that runs an AI reviewer on all changes - checking for security issues, test coverage, style, and complexity - before any human sees the PR. Only escalate nuanced decisions.
  • Why Now?: High-velocity AI-generated code is flooding repositories, and manual review doesn't scale. Models now reliably catch technical debt and policy violations, especially when guided by standardized prompts.
  • Expected Upside: 50 - 80% reduction in human review time, improved code consistency, and earlier catching of critical issues - enabling you to maintain quality while shipping faster as a solo developer.

• Adopt a model-driven architecture by minimizing custom orchestration logic and letting modern AI models handle core reasoning and tool usage, reducing development time and maintenance overhead.
• Build or use declarative, developer-friendly SDKs (like Strands Agents) that focus on essential decisions - model choice, system prompt, and tools - to accelerate agent development and deployment.
• Simplify agent design by removing unnecessary layers such as multi-agent loops, streaming APIs, or complex state management, especially when targeting HTTP-based or serverless environments.
• Implement AI-powered code reviews before human review to catch formatting, security, and maintainability issues automatically, ensuring higher-quality submissions and freeing developers for strategic work.
• Standardize and share reusable skills or API integrations through centralized, composable systems (e.g., a single MCP server with team-owned skills) instead of allowing fragmented, siloed agent implementations across projects.