In this KiwiSaaS webinar, Tony Schollum, CTO of Re-Leased, shares how the company went all-in on agentic engineering — and what changed once AI accelerated the process of writing code. Tony offers a practical look at the team’s rollout, the new bottlenecks that emerged, and how Re-Leased is redesigning engineering workflows around agents, evidence and human judgement.
Key themes:
1. Move beyond individual AI experimentation
Re-Leased found that giving engineers time and budget to experiment with AI tools created uneven results: some people moved quickly, while others fell behind. The team responded by making agentic engineering a shared way of working, rather than an optional pilot.
Its initial model centred on a reflective loop: plan, build, review, reflect. Instead of manually correcting an agent’s output, the team worked to understand why it went wrong and improved the surrounding instructions, skills, memory files and documentation. As Tony puts it: “When the agent gets it wrong, you don’t fix the code, you fix the instructions.”
2. Build harnesses that verify the work
Re-Leased built an internal harness called Reflow to run repeatable engineering pipelines for bugs, features and migrations. These pipelines can reproduce a bug, plan and implement a change, run checks, document testing and create a draft pull request.
The goal is not simply to produce code quickly. It is to produce a fully tested pull request with evidence that the work meets the relevant acceptance criteria. That evidence can include screenshots, descriptions and a narrated video of the testing completed.
3. Use deterministic checks wherever possible
A major learning was to separate the work that needs AI reasoning from checks that are clear-cut. Re-Leased uses deterministic tools such as linters, unit tests and integration tests within its pipelines rather than asking an AI model to handle every verification step.
Tony’s three principles for a good harness are to recreate the steps and feedback loops people already use, make as many gates as possible deterministic, and require evidence that the job is done.
4. Shift quality and review to where they add most value
Once code generation sped up, Re-Leased found that refinement, testing and review became the new constraints. Engineers were spending more than 30% of their capacity reviewing pull requests, while QA was overwhelmed by the volume of change.
The response has been to move testing earlier. Engineers now own testing through to production, while QA has shifted towards quality advisory work: helping shape test plans, supporting exploratory testing and strengthening automation.
Re-Leased is also trialling security and risk-scoring agents in its pull-request process, with the aim of focusing human attention on work with the greatest potential impact.
5. Don’t let agents amplify existing drift
Agentic engineering can expose and accelerate existing weaknesses in a codebase. At Re-Leased, the team found UI drift was happening four times faster after introducing AI coding tools, as agents copied nearby examples rather than consistently following the design system.
The lesson was not to blame the tool. It was to improve the system around it: the instructions, component libraries, themes and examples that agents draw from. As Tony says: “The tool is not the problem, it’s what we gave it.”
6. Keep humans focused on the novel work
Tony is clear that this approach is not about replacing engineers. It is about reducing repetitive work so people can focus on the work that needs deeper judgement: new architecture, novel product problems and complex customer needs.
That also means changing how product and engineering collaborate. Rather than relying only on polished handovers, Re-Leased is bringing engineers closer to raw customer context, including interviews, transcripts and prototypes.
Key takeaways:
Make agentic engineering a team capability. Move beyond optional, individual experimentation and give teams practical support to adopt a shared way of working.
Start with one repeatable workflow. Build a small harness that includes a verification loop, rather than trying to automate everything at once.
Make evidence a requirement. Expect every agent-led workflow to show how the work was tested and why it is ready for review.
Use AI and people for the work they do best. Automate repeatable checks and lower-value toil; focus human judgement on higher-risk, ambiguous and novel work.
Keep measuring for the next constraint. Faster code generation shifts pressure elsewhere, including testing, review, handovers and design-system consistency.
Watch the full discussion to explore how Re-Leased is putting agentic engineering into practice and what it is learning along the way.