Agentic delivery without the vibe coding

Result: in the teams I work with, the tasks each engineer completes per two-week sprint rose from about one to as many as three or four, while every regular and ad-hoc meeting stayed in the schedule. Quality is protected by human review at every step.

Context

Two enterprise programmes in the insurance and automotive-claims sector. The current one is a policy-management system for an insurer operating in the EU and the US, delivered by a team of nine. Earlier, I spent two months on a parts-and-repairs supply platform inside a shared claims-management ecosystem. There, the engineering development manager chose AWS's AI-driven development lifecycle framework (AI-DLC), and I applied it as the only engineer in a sub-team of a three-to-four-team programme.

The problem

AI coding tools were already allowed, but for most of the team agentic development was a new idea. Used casually, "vibe coding" produces code nobody fully understands, decisions nobody wrote down and changes nobody tested. The goal was the speed of AI without those side effects.

What I built

A Development Orchestrator agent. It follows ten to eleven steps, each written down as a skill: a detailed, reusable instruction set the agent must follow. Given a Jira ticket number, it:

1. gathers the task context;
2. reproduces the problem first, if the ticket is a bug;
3. writes the confirmation test cases before any implementation;
4. implements the change;
5. opens the pull request and handles review comments;
6. builds the Docker images;
7. automates verification with Playwright and Postman.

A Refinement agent. It finds the blocking questions for the Product Owner, analyses the scope of a change (how many services and repositories it touches) and presents that as an easy-to-read infographic so the meeting moves faster. The meeting recording is then used by the same agent to write a summary for every task and store it in the task itself, as context for future work.

How it stays safe

• A checkpoint at every step. The engineer reviews the plan and the research reports before the agent moves on.
• Tests before code. Bugs are reproduced and confirmation tests are written first.
• The process is written down. Each step is a skill, so it is repeatable and open to review.
• Personal data never reaches the model. Insurance and personal data (PII) is redacted in code before anything is sent to an AI service.
• Context is kept. Every task carries its own history, so knowledge is not lost between meetings.

What changed

• Throughput: tasks completed per engineer per two-week sprint went from about one to as many as three or four, with all standard and ad-hoc meetings still taking place.
• Standard SDLC meetings, refinement in particular, run faster.
• Documentation, design and architecture decisions are generated and kept up to date.
• Features are implemented and iterated on faster, with tests in place from the start.

Sharing it

I built these agents on my own initiative for my daily work and made them configurable, so other teams can reuse them. For colleagues, agentic development was an entirely new concept, so I help them understand it in practice, including pairing sessions.

What I learned

The speed comes from the model; the safety has to come from the system around it. Building an approach that protects a team from the downsides of vibe coding matters more than any single prompt.

What this means for your team

I can help you with three things:

• Process assessment: where your delivery process loses time and where AI can help without adding risk.
• Introducing and improving AI-assisted workflows: agents with human checkpoints, tests first and privacy built in.
• Training: hands-on sessions that help your engineers work with agentic tools confidently.

Talk to me