AWS's Marc Brooker: Specs, Not Code, Are the Hard Part

The podcast discusses the evolving role of AI and automation in software development, particularly in on-call engineering and incident management. Automation tools, including AI agents, are increasingly handling routine tasks such as log analysis and metric monitoring, allowing human engineers to focus on complex problem-solving, system resilience, and diagnosing large-scale or metastable failures. Insights from thousands of post-mortems highlight the importance of humility, system design, and the limitations of AI in understanding ambiguous or subjective issues. While AI can preprocess incidents and learn from past failures through semantic search and procedural memory, human judgment remains essential for nuanced decision-making and long-term improvements.

A significant focus is on the development and deployment of AI agents, with an emphasis on trust, policy, and security. Tools like AWS Strands and policy languages such as Dogwood aim to create deterministic, secure frameworks for agent behavior, enabling autonomy while maintaining control. The discussion also explores how specifications, testing, and oracles are becoming central to software development, as defining clear requirements becomes more critical than coding itself. Additionally, the industry is shifting toward leveraging AI agents for tasks like design reviews and incident response, while recognizing the need for mentorship, knowledge transfer, and integrating AI-native developers who bring fresh, innovative approaches to engineering practices.

• What if you built a personal AI agent to handle your software business's routine monitoring and incident triage?

  • Move: Use AWS Strands SDK to build a lightweight agent in the next 3 days that ingests logs from your production app, runs basic anomaly detection (e.g., error rate spikes), and posts alerts to a Slack channel with suggested runbook steps.
  • Why Now?: On-call fatigue slows down solo development; automating first-response triage frees you for higher-leverage work, and AWS Strands lowers the barrier to agent creation with ready-made tool integrations.
  • Expected Upside: Reduce incident response time by 50% and reclaim 3 - 5 hours/week currently spent on alert investigation, enabling faster iteration on core product features.

• What if you trained your AI agent to learn from every bug fix and postmortem in your codebase?

  • Move: Over the next two weeks, create a structured log of every production issue (including root cause and resolution), then set up semantic search over these logs so your agent can retrieve and apply past fixes during new incidents.
  • Why Now?: You're already writing postmortems informally - formalizing them now creates compound value as AI tools evolve, and retrieval-augmented generation (RAG) makes this actionable without complex AI training.
  • Expected Upside: Cut debugging time for recurring issues by 60% and reduce repeat outages, turning operational experience into an automated knowledge advantage.

• What if you offloaded specification writing to an AI agent guided by your testing oracles and customer feedback?

  • Move: In the next 5 days, define a small feature (e.g., a user onboarding flow) using only test cases and acceptance criteria, then prompt an AI agent to generate a formal specification and corresponding implementation plan.
  • Why Now?: The bottleneck in solo development is often unclear requirements - not coding - and AI excels at structuring ambiguous inputs when anchored to real feedback signals like user behavior or test outcomes.
  • Expected Upside: Shorten feature definition time from days to hours and increase alignment with user needs, accelerating your build-measure-learn loop.

• Implement a basic AI agent using AWS Strands SDK within the next week to gain hands-on understanding of agent mechanics and tool integrations.
• Create an agents.md file to define clear expectations and instructions for any AI agents used in your development workflow, ensuring alignment with desired outcomes.
• Analyze past project post-mortems (or simulate them) to identify recurring failure patterns and encode those learnings into automated checks or agent decision rules.
• Focus on writing crisp, deterministic specifications for new features or systems before coding, treating tests as executable specifications to guide AI-assisted development.
• Set up a personal policy framework (e.g., using natural language rules) to govern AI agent behavior - such as data access or code modification boundaries - to improve security and reliability in automated workflows.