Starting is no longer the problem
Most people using coding agents have a good system for starting a project. They do not yet have one that keeps them going.
I have been speaking with a handful of “new developers”. These are non-native software engineers who are actively building products. They can prompt an agent, connect to a database (often Supabase), and deploy a prototype in a matter of days. They are being empowered by their organizations to continue their development journey often because they understand how the business operates from years of experience.
The challenge these developers face is that as their codebase continues to grow, every new decision has a larger blast radius. They often open Claude or Codex at the beginning of their day and reconstruct where they left off. What decisions did we make? What should we do next? Is this change safe?
The risk & security surface increases as they continue to prompt the codebase into existence.
The minimum loop
New developers have a setup that is designed for building a prototype, not for continuously improving a product in production. As more people start building their own software, I think they will need a minimum viable loop. The loop needs three things:
- Durable context. Any project needs a machine-readable context store. This is a space where the project plan, itemized issues, and a register of important decisions can be easily called by an agent. The previous session logs should not be the only place where the project makes sense. If a developer has to do a sensitive task (such as a database migration), they should build a runbook that any agent can read to walk them safely through the task.
- Automatic feedback. Code getting pushed to source control should always go through a series of gates. A test suite, preview deployment, and code review should trigger with most every commit. The goal is not to remove human judgement from merging new code; it is to ensure there is a quality control process that acts as a first layer of verification in every development workflow.
- A clean handoff. Every agent session leaves residual context. At the end of each session, the context store should be updated. This can be as simple as closing an open ticket in Linear, adding comments about what work was done, and noting the next action or ticket that should be tackled.
A good loop should make it possible for a team to start any new day with clear context on what was completed and what needs to be done. It should make it easy to pause work and return to safely complete a task. And it should employ an agent to do a bounded task with a clear goal and stopping point.
From individual habit to team system
The larger loop connects the human, the agent, and the project across many sessions. This becomes even more important with a team. At the minimum, different build styles create drift. At the maximum, an organization expands its surface area for security risk without knowing what has been deployed or who owns it. This is how AI adoption begins to outpace governance.
