A software factory needs a repeatable process.
For me, a software factory means a way to move from an idea to working software repeatedly: scope the work, make decisions explicit, give agents useful context, and build review into the process.
I want speed to come from better coordination and less repeated work. Clean boundaries, typed contracts, and tests make it easier to change a system while keeping the rest of it working.
Autonomy needs an operating environment.
I want agents to carry work forward on their own when the task, tools, rules, and system limits are defined. That includes scheduled work, access to the right apps, durable context, and a clear route for exceptions. A person should be able to set the boundaries and inspect what happened.
Coordinating the development environment.
I work with 10+ Claude Max and Codex subscriptions, alongside Kimi Code. Neomax, my Rust orchestrator, coordinates providers, models, accounts, sessions, and worktrees. Managed worker tasks preserve their state through interruptions and keep routing within the allowed scope.
Across Neotask, I use repeatable workflows for changes that touch several repositories. Adding an app provider means checking authentication, tool execution, interface behavior, and the public documentation together. An upstream release needs its own baseline, change record, verification, and recovery path.
Swarms need ownership.
Several agents can work on independent problems at the same time. They still need clear responsibilities, a shared understanding of the interfaces, and someone responsible for integrating the result.
That is part of what I’m exploring through Neotask and my development workflow: how to coordinate agents without losing context, control, or the ability to inspect what happened.
Tools should help the work compound.
I use tools based on and inspired by Jeffrey Emanuel’s Agentic Coding Flywheel for coordination, task tracking, session search, and memory. His work is an upstream foundation I use and build on, with credit to the people behind it.
The principle is practical: make past decisions easier to find, make current work easier to coordinate, and make mistakes easier to catch.
The result has to survive use.
I’ve built systems that collected live market data, ran purchasing workflows, and supported content enforcement. Those projects taught me to inspect the complete path, including the parts that fail when conditions change.
With AI, I bring the same approach to the work: examine assumptions, verify behavior, and keep people involved in the decisions that deserve their judgment.