Eno Reyes, co-founder and CTO of Factory, presents how to build a "software factory" — an organization where agents take on the full software development loop. He starts from a retrospective: in 2023, code AI was expected to progress from line autocompletion to file autocompletion, then to repositories. Coding agents arrived, but most usage remains "I ask, I get." According to him, this era is running out of steam: organizations are now asking questions about cost overruns, return on investment, dependence on a single model, and the role of developers, who spend little time on coding alone.
He describes development as a feedback loop. Signals coming from customers, bugs, telemetry, or competitors are triaged, then turned into a plan, code, validation, and deployment, which in turn produces new signals. Each step can be handed to agents. For Reyes, the real difficulties are measurement, analytics, and governance: the problem is organizational rather than technical.
Three principles structure his framework. First, model independence: an IT leader should not have to tie themselves to a single vendor, nor be subject to its pricing. A routeur de modèles allocates compute by task and by role. Second, sovereignty: choosing one's deployment, from SaaS to air gap, and retaining control of traces, since correcting a bad decision made by the factory requires understanding why it was made. Third, coverage of the entire lifecycle, with the same agent across all surfaces, to avoid scattering policies, routers, and trace exports.
Governance remains difficult: agent identity (inherited from the user for interactive activity, service account for sandboxed review), model controls, allow lists, risk classification, spending caps, and audit capable of tracing, in a single query, from table access back to the person, the prompt, and the traces. He considers indexing code in a cloud unnecessary.
The bottleneck, he says, is no longer the model, nor now the harness, but the organization's preparation: tools usable by agents and deterministic validation (lint, tests, typing, conventions). Factory offers an agent readiness product measuring more than 150 signals.
Humans become the stewards of the factory: they watch for drift, correct the system, and durably prevent errors, in regular code rather than natural-language specifications. Reyes concludes that companies will need to find their "Google metric," a link between an engineering indicator and a business outcome, to allocate inference according to ROI, for example by giving an agent a performance target with a token budget.