@chenchengpro: Everyone's talking about the agent's "loop" lately, but few explain what it actually is. Warp CEO Zach Lloyd gave a practical version: a two-loop mechanism where Skills self-evolve from feedback, using GitHub issue triage as an example. Inner loop: each new…

Factory 2 seems to have already done a great job engineering the Loop.

Factory founder Matan Grinberg announced Factory 2.0, setting the tone in one sentence: improving individual engineer productivity is no longer enough. What truly needs to be unlocked is organizational productivity, and it doesn’t require faster code completion — it needs an interconnected, agent-native, end-to-end system that continuously improves by observing itself, where the minimum increment unit is an AI agent. He calls this system the “Software Factory.”

The Software Factory is less a new tool and more a system topology: starting from external signals (bug reports, internal conversations, customer feedback, business requirements) → triaged into planned changes → built/tested/reviewed/hardened/released/monitored → monitoring then generates new signals, forming a continuous feedback loop where the loop itself is the product. The author’s assessment is that almost no one has truly made this loop fully AI-driven yet; it’s still early, but the spread will be fast.

It sets three hard pillars for a “robust Software Factory.” First is Model Independence: no single model fits all enterprise needs; you must deliberately select models for different tasks, or use a Router to automatically or rule-based pick the “best” model based on cost/performance/speed, hedging against cost and capability changes from model commoditization. Second is Sovereign Intelligence: deployment forms range from fully managed cloud, bring-your-own-key, self-hosted data plane, EU-specific, all the way to fully air-gapped with no internet; but the focus of sovereignty isn’t “where it runs” but in owning a system that learns from itself — every agent session, code review, resolved incident flows back into the loop, keeping capabilities forever within your walls. Third is Continual Learning: every stage of the SDLC must be instrumented — code review/security analysis/documentation/QA/incident response run on the same platform, sharing the same agent core, router, and organizational context — so security findings can feed back into code review, deployments can trigger documentation updates, incidents can be linked back to the PR that caused them.

These aren’t just concepts; they are already running in production environments at organizations like NVIDIA, EY, Adobe, Palo Alto Networks, Adyen, Blackstone, Wipro, Comarch, and others. Autonomy is treated as a spectrum, not a switch: well-defined tasks use simple Droid agents or skills, periodic workflows use Automations with shared goals and memory, remote persistent execution uses Droid Computers, complex tasks use Missions to split work into parallel tracks spanning hours or days. Finally, landing on the “human” aspect: engineers are no longer the sole guardians of building software, but instead build the factory that builds software, thereby taking ownership of governance, security, and business outcomes — the next era is engineering-led.