AI coding assistants are now standard in every tech stack. They work well for individual users, but they hit a wall when teams scale. In a recent presentation, SumZero developer Maarten Anckaert explained how his agency replaced standard AI tooling with a fully custom development harness. The result was lower API costs, better context management and higher developer output, with human engineers still firmly in control.
Standard copilots look magical in a demo. After a few hours of real commercial work, their limits become obvious. Default assistants quietly load huge amounts of background context with almost every query, even a simple prompt. This bloat slows everything down. Worse, once a model fills more than about 55% of its context window, it starts hallucinating code that doesn't exist, and every mistake burns more expensive tokens.
SumZero first experimented with markdown prompt agents, custom editor extensions and open-source toolkits like pi.dev. Those tests proved the value of owning the AI pipeline. Maintaining bespoke tooling on a hundred developer laptops, however, became an administrative headache. The team needed one central platform that could do the heavy lifting and protect client data.
The result is a scalable harness with thirty specialized agents and more than eighty technical skills. At its center sits a lightweight orchestrator model that routes each request to dedicated planner, developer or reviewer agents. Developers no longer debate which model suits which task. This single change cut AI operating costs to just 5% of the previous level.
To retain knowledge, the team built Hermes memory, a local memory engine that tracks developer feedback and code quality standards over time. Agents automatically check these lessons when generating new code. Because everything runs locally, sensitive client code never leaves the building.
The third pillar is deterministic output. Rather than prompting an AI for the same task over and over, use it once to write a script. Maarten had an AI write a Python script that generates slide decks in two seconds, with identical results every time and no further prompts. The same logic applies to browser testing: let AI write the test once, then simply run it.
During the Q&A, our audience asked how to keep sixty local setups in sync on a shared codebase. For now, local environments keep code safe and stable. The next step is securely pooling that learned knowledge across the organization, so a new hire instantly inherits the workflows, rules and fixes from hundreds of past projects.
This efficiency also changes the business model. Under time-and-materials billing, faster teams earn less because they bill fewer hours. SumZero is therefore moving toward fixed-price projects, letting clients benefit from faster delivery while the agency benefits from its platform investment.
Off-the-shelf copilots only go so far in an enterprise setting. By combining smart orchestration, persistent local memory and script-based automation, teams can control their tooling, manage context, keep code quality high and cut compute costs. As the technology matures, the real competitive edge will come from capturing team knowledge and shipping reliable, predictable software.
%20copy.webp)
