AWS Veteran: How The New Software Development Life Cycle Works

In plain English

Heitor spent 11 years at AWS and now works on a resilience team at the payments company Adyen. He walks through how his team builds software with AI agents (software that carries out tasks on its own), step by step.

People still talk to customers and sketch the plan on a whiteboard before any AI is involved. Then the plan becomes a written roadmap with clear goals, a detailed design, and small tasks the AI carries out. Checks at the end confirm the work actually met the goal.

His big lesson is that the preparation and checking matter far more than the typing, and that you should never simply trust an agent’s word that it did the work.

Three ideas worth keeping

• Use the most capable AI for planning, a mid-range one for building and cheaper ones for repeated reviews, because costs add up fast across a large team.
• Agents sometimes claim they ran tests they never ran, so his team builds checks that record proof that each review really happened.
• After each session, run a short “retro”: the AI interviews you about where you had to correct it, then lists which fixes could become automatic rules.

What it means for your business

AI does not remove the need for people who understand the customer and the problem. It moves their effort to the start (deciding what to build) and the end (checking it worked). Expect AI spend per engineer to become a real budget line, and ask your team how they keep it sensible.

The best policy guides people toward good habits rather than banning everything, because bans just push people to unofficial tools.

What it means for software teams

Start with a simple spec-driven tool, but don’t stop at its defaults: build your own team’s steps, checks and review stages on top. Clear the AI’s working memory between planning and building, since the plan is already written down. Run reviewers matched to what changed, with a few that always run.

What we found, and what we did

We do our own version of his retro. We went back through three days of agent transcripts in June (4,439 messages across 13 sessions) and every session from a bad week in July.

We turned the failure patterns into fixes, written rules for our pipeline and guidelines for our AI reviewer. The cycle behind it is in Compounding Engineering: The 4-Step Loop.

We also caught agents marking changes “audited” without running the audit. We replaced that honour system with a step that actually runs the audit and records the result.

Where we differ is on piling up AI reviewers. In our own blind test the best setup caught only 4 of 11 known defects. So our AI reviewer advises, and the hard decisions rest on automated checks and, for anything hard to undo, one person. More in Our Journey to Automated Code Review.

Worth watching

• 56:39: the three tiers of AI models and why cost forces the choice.
• 1:36:44: the “merge check” that stops agents faking evidence of reviews.
• 1:38:36: the retro command that turns each session’s corrections into new rules.