We're IBM Research's proving ground for emerging AI.

A human-operated software factory, not a dark one

Our take on what a software factory is, what it means for the business of software engineering, and how someone builds one on DAM. Two halves: the vision, then the opportunity. The vision the team is arguing for is a **human-operated factory, not a dark one**. The claim is not that software development gets automated away — it is that the collaboration between the person and the agent gets automated, as far as it will go, and that the models are not good enough for anything else today. That makes the interesting question *where* the person belongs rather than whether they do. The answer the team keeps arriving at is a small set of defined touch points — context creation, business and architecture alignment, review and approval — with everything between them automated. Two kinds of alignment do different jobs and should not be collapsed: a technical grill with the agent surfaces unknown problems, a discussion with people creates agreement on direction. The structural argument is that a factory is a **codified process, not a group of agents deciding together**. Implement, review, approve, merge — fixed, which is what makes parallel work safe; freely coordinating agents produce race conditions instead. GitHub is both the channel and the state store, so the whole factory state lives in issues, labels, PRs and commits: stop it, start it again, and it continues without duplicating work, and people join in the same place the agents work. Agents hand off through the channel rather than to each other directly, because direct handover has no confirmation, no shared state and bad timing. Events push work to agents; agents do not wake on a timer and go looking for it, which mostly buys wake-ups that find nothing. The second half is where DAM comes in — which building blocks exist, which are missing, and what has to be added before someone can build their own factory on it. The framing to keep: building blocks are not the story. Start from the audience, then the pain, then the features, then the blocks. And the team is its own first customer — building our own factory on DAM is how we find out what is missing. The post has to take positions against the state of the art rather than describe it: factory.ai, the KubeStellar hive, and how Anthropic build and run their own factories — including that they started with free agent groups and moved away from them. Where those disagree with us, say so and say why.