Developing an enterprise-level AI platform for the world’s largest marketing organizations requires our engineering team to move quickly while meeting industry-leading expectations for quality, compliance and data protection. Those expectations also shape how we use AI to get the most value out of our engineers while leveraging agents to automate manual work.  

As agents take on more implementation and repetitive work, our engineers spend more time defining problems, planning solutions, and refining what gets built. This is the first of two posts on our approach to agentic workflows. Here, we explain how we divide the work and how that division is changing our development practices. 

AI can't buy us taste 

Our starting principle is that AI only adds value when it strengthens our engineering judgement. Producing more code is only useful if we’re building the right thing and can stand behind the result. We divide the work into three categories, according to how much creativity and accountability each task requires: 

Autonomous: Engineering teams have always automated repetitive tasks. AI expands the range of work we can automate and reduces the effort required to do it. At Bluefish, examples include initial incident investigation, standard build fixes, first pass code review, and dashboard creation. Each has a defined scope that gives an agent a clear task to execute. 

Human-in-the-loop: Some workflows need a person to interpret results and decide what happens next. Our evaluation pipeline for AI features runs test sets, uses AI to grade the output, and flags results below our quality targets. An engineer then decides how to best tune that pipeline. 

Human-led: Picking the right data set and designing campaigns to create the best marketing outcome is where humans excel. These decisions require marketers to understand the customer’s goal, weigh tradeoffs, and take responsibility for the choice.  

These categories are core to our approach of establishing who owns each decision. AI executes well inside clear constraints, but can’t produce original thought, so the accountability stays with our engineers. Our engineers define what good looks like, and AI works to that standard. As AI produces more output, that judgement becomes even more important. 

Pair planning replaces pair programming 

We’re constantly iterating how AI fits into our engineering stack, but one trend is clear. About 90% of Bluefish engineering time now goes to planning and code review. 

Most teams now follow spec-driven development, defining the intended behavior and implementation approach before handing off work to agents. Some have replaced pair programming with pair planning, where two engineers decide together how to break a task down for AI agents. Once that plan is established, engineers focus on breaking down the plan into logical steps for AI agents to take on. 

Review practices are evolving alongside implementation. We are building always-on guardrails for scanning code, designed to ensure that the latest standards are applied. Our most senior engineers are also constantly updating these guidelines and guardrails. Increasingly, software engineering is becoming more focused on building the harness and guardrails around workflows. 

The biggest change is outside engineering 

An even bigger shift is that for the first time, our business teams can build software and automations themselves. Our go-to-market teams are among Bluefish’s heaviest AI users. That gives more people the ability to solve problems directly, while introducing new ownership questions that engineering teams have long had to address: Who owns the build pipeline, who fixes it when something breaks, and how does knowledge survive beyond the person who created a tool? 

We’ve found success distributing skills through internal marketplaces for engineering and non-technical teams. Each skill packages a best practice with the tools and guardrails needed to apply it, so results stay consistent. Every skill passes automated checks for secrets, personal data, and vulnerabilities before it is published, and periodic audits re-scan what is already live. This lowers the risk of sharing automations across the company.  

Deployment is the next problem to solve. We need to give non-engineers a secure way to deploy what they build, with clear maintained responsibilities defined once others depend on it. 

What comes next 

We're iterating every day to lead in AI native development workflows. Our focus is on what we can control: we invest to keep our engineers at the forefront of new tools and practices, so they become more effective and more satisfied in their work.  

In part two, we will explain the metrics we track, from billed cost to delivery results, and the guardrails that keep quality high as we scale.