How we work
How we work
High output with AI agents. Every step reviewed.
An AI agent can write an astonishing amount of code in a short time. But it has a problem reminiscent of the main character in the film Memento: it cannot permanently remember anything new. Whatever does not fit into its current context is gone, at the latest with the next session.
That is why we give our agents a memory in files: a plan with verifiable criteria, progress notes for every step and a handover for the next start. Forgetting is planned for, and the agent still reaches its goal.
This website is built the same way
We are building this website with exactly this process: an agent implements it step by step, a second model reviews every step and rejects it when something is wrong, and we as people sign off design and content before they go live.
Our process
We develop in a fixed process: a second, independent model reviews every step against the plan before work continues.
Plan before code
Goal, scope and acceptance criteria are fixed before an agent starts.
Small steps
Implementation with tests, version control and documented progress.
Independent curator
A second model reviews every step and blocks it when something is wrong.
Humans decide
A passing build does not replace a review by people who know the subject.
What carries the process
Memory in files
Plan, progress, handover and project knowledge live in files inside the project. Every new session starts from the real state, not from memories.
A second look
The curator is a different model that may only read. It judges against the acceptance criteria and can block a step.
Clear limits
After a fixed number of rounds without approval, the process stops. Then a person decides, not the agent.
Local if needed
Agents can work with local models when code or data must not leave the building.
Does this fit your task?
Tell us about your use case and your data. We will get back to you personally.
info@masterbrain-bytes.com