"Software Factories" was the theme of AI Engineer World's Fair 2026, held in San Francisco in late June and early July 2026 — it refers to an industry consensus taking shape around "handing the entire software development lifecycle to a self-running loop of AI agents." The conference's keynote lineup brought together Microsoft, OpenAI, the open-source camp represented by Z.ai (GLM) and MiniMax, and Hugging Face, each interpreting what a "software factory" should look like in their own product language. On the surface, everyone was saying the same thing — "use AI agents to automate the development process" — but pull apart each company's actual argument and you'll find they're standing in genuinely different places on questions like "who's responsible for what inside the factory" and "open source or closed." This piece doesn't try to declare a winner; it lays out the key points of disagreement so enterprises and developers evaluating AI development-agent platforms can see clearly what they're actually choosing between.

Where This Debate Starts: What AI Engineer World's Fair 2026 and "Software Factories" Actually Are

AI Engineer World's Fair 2026 ran June 29 to July 2, 2026 at Moscone West in San Francisco — an industry conference focused on AI engineering and applied deployment, with this year's theme set as "Software Factories." According to a definition offered in a talk by Tereza Tížková, representing conference co-curator Factory, a software factory means "a self-running loop across the entire software development lifecycle" — concretely, gathering signals, reacting to user feedback and system logs, prioritizing, and coordinating integration. Her core idea: let AI agents do the actual work of "building," while the developer's role shifts up a level, into designing and maintaining the system that lets agents write code on their own. That framing is the shared backdrop against which every company then staked out its own position.

Point of Disagreement One: Is a Software Factory an "Agent Loop" or a "Productized Platform"? Microsoft and OpenAI Speak Two Different Languages

At the same conference, both Microsoft and OpenAI were saying "AI agents make development more efficient" — but using two different language frameworks.

Microsoft was represented by Pablo Castro, introducing Microsoft Foundry and positioning it as an "AI app and agent factory." His core argument was that human-agent collaboration produces a "learning loop" — in other words, Microsoft's talk emphasized a platformed "factory" concept that houses agents and applications, with the focus on how the human-machine collaboration process itself keeps accumulating and improving.

OpenAI was represented by Alexander Embiricos and Romain Huet, discussing how Codex agents boost productivity through stacked agent loops. Their core argument: if you can connect an agent to "what the work is" and "why it's being done," and then to the subsequent review and deployment process, the agent can genuinely get more work done — in other words, OpenAI's emphasis is on the agent loop itself connecting to the front and back ends of the development process (why this work needs doing, who reviews it once it's done, how it gets deployed), rather than just automating the code-generation step. The conference's opening keynote, meanwhile, was given by AI Engineer World's Fair/Latent Space cofounder swyx (Shawn Wang), titled "Loopcraft: The Art of Stacking Loops" — he was speaking as a conference organizer, not an OpenAI spokesperson, but the talk likewise centered on "loops," tracing how the form of developer-AI interaction has evolved from simple chat, to using tools, to setting goals, and on to automation and loop design becoming a craft in itself.

Put these two companies' arguments side by side, and the difference isn't about whose technology is stronger — it's where each narrative enters: Microsoft puts the emphasis on the "factory," the platform that houses everything; OpenAI puts the emphasis on the "loop" itself — how it stacks, how it connects to the front and back end of the process.

Point of Disagreement Two: How Does the Open-Source Camp See It? Z.ai's GLM-5.2 and MiniMax's M3

If Microsoft and OpenAI represent the closed-source, platformed route, the other side of the conference had representatives from the open-source model camp making their case.

Z.ai (the GLM team) presented remotely via Zixuan Li, launching the open-source large language model GLM-5.2, positioned as a "flagship model for long-horizon tasks." Z.ai also released a tool called ZCode, claimed to "support all frontier models," and compared ZCode to OpenAI Codex in the talk. That signals the route Z.ai has chosen: pairing an open-source model with a tool layer that claims compatibility across multiple frontier models, rather than binding the agent tooling to its own model inside a single closed ecosystem the way Microsoft or OpenAI does.

MiniMax, the Chinese company, was represented by Olive Song, who announced its latest open-weight model, M3, in an interview conducted by Hugging Face's Thomas Wolf. Public reporting so far hasn't gone into much detail on M3's specific technical characteristics, and this piece won't speculate beyond what's been reported — the confirmable fact is limited to this frame: "MiniMax representative Olive Song announced M3, an open-weight model, at this conference, and was interviewed by Hugging Face's Thomas Wolf." Performance figures, training scale, and similar details still await further public information.

Worth noting: GLM-5.2 and M3 belong to two different companies (Z.ai and MiniMax) and shouldn't be conflated as the same product on the same route; the two share only that they're both released as open-weight/open-source models — which, set against Microsoft's and OpenAI's relatively closed-source, productized routes, forms a second, parallel narrative thread running through the conference.

Point of Disagreement Three: Who Gets to Define "Software Factory"?

One easily overlooked but worth-noting point of contention: the definitional framing behind the conference theme itself, "Software Factories," came from Factory's Tížková, not from Microsoft, OpenAI, or any of the open-source camp companies. In other words, when each company used words like "factory" and "loop" in their talks, they were borrowing or echoing the same conference-level narrative framework and filling it in with their own product language — but the right to originally define that framework belonged to one of the event's co-curators. That's also why different companies at the same conference can end up with genuinely different concrete interpretations of "software factory" — because this is a concept still being collectively defined, not a technical spec with one agreed-upon answer yet.

Both Sides Are Actually Saying the Same Thing: Looping and Automation Are the Consensus — the Disagreement Is Over Who Controls the Loop

Put these points of disagreement together and a neutral observation emerges: whether it's Microsoft's "learning loop," the stacked-agent-loop approach OpenAI emphasizes, or Z.ai's long-horizon-task model, nearly every company at this conference agreed on the broad direction that "the development process is moving toward a self-running loop" — that's the consensus, not the disagreement. The real divide is over where that loop runs, whose model drives it, and whether developers can swap out any given piece of it. Microsoft and OpenAI tend to present the loop as bundled with their own platform and their own model; Z.ai, through its "ZCode supports all frontier models" positioning, emphasizes that the tool layer and the model layer can be chosen separately. This is a structural difference in approach, not a comparison of which technology is better — enterprises evaluating these platforms should treat the two as separate questions.

What This Means for Enterprises and Developers: What Metrics to Look at When Evaluating an AI Development-Agent Platform

Turning this debate's points of disagreement into a practical checklist for enterprises evaluating AI development-agent platforms, there are roughly three directions. First: is the platform's "loop" design locked to a single model vendor, or does it allow swapping in different models at different steps — this directly affects your future negotiating room and vendor-dependency risk. Second: how far does the platform's "automation" actually go — just code generation, or does it, as OpenAI describes, connect through to why the work is being done, review, and deployment as a full pipeline; different depths of automation imply very different integration complexity and governance needs. Third: open source and closed source aren't inherently good or bad, but they represent different operational assumptions — an open-source model gives an enterprise more flexibility to self-host and customize, but requires taking on more operational responsibility itself; a closed platform absorbs that operational complexity, but the enterprise has relatively less control over the underlying model. These three metrics are, at bottom, the same logic an enterprise should be asking when choosing any multi-model access solution.

FAQ

What is "Software Factories," and why did it become the theme of this 2026 AI conference?

"Software Factories" was the theme of AI Engineer World's Fair 2026. According to the definition offered by Factory's Tereza Tížková, it means "a self-running loop across the entire software development lifecycle," encompassing gathering signals, reacting to feedback and logs, prioritizing, and coordinating integration; the core idea is letting AI agents do the actual building work, while developers shift into building the automated system itself.

Is Microsoft's "learning loop" the same thing as OpenAI's stacked-agent-loop approach?

Not entirely. Microsoft's Foundry, introduced by Pablo Castro, is positioned as an "AI app and agent factory," with the core argument that human-agent collaboration produces a learning loop; OpenAI's Codex agent, discussed by Alexander Embiricos and Romain Huet, emphasizes stacking multiple agent loops and connecting an agent to what the work is, why it's being done, and the review and deployment process. Both describe an automated development process using the word "loop," but they enter from different angles and emphasize different things — they shouldn't be treated as the same design.

What's the difference between Z.ai's GLM-5.2 and the OpenAI Codex approach?

GLM-5.2, released by Z.ai, is an open-source large language model positioned as a flagship model for long-horizon tasks, paired with a tool called ZCode that claims to support all frontier models; OpenAI Codex is an agent product bound to OpenAI's own model and platform ecosystem. The difference is open source vs. closed source, and whether the tool layer is tied to a specific model — this piece doesn't judge which approach is technically better.

What specific technical specs did MiniMax announce for M3?

Public reporting so far hasn't gone into much detail on M3's specific technical characteristics; the only confirmable fact is that MiniMax representative Olive Song announced this open-weight model at the conference and was interviewed by Hugging Face's Thomas Wolf. This piece won't speculate on unverified performance figures or training scale — watch for official announcements from MiniMax.

Should enterprises prioritize open-source or closed-source routes when evaluating AI development-agent platforms?

The conference presented two parallel routes, with no absolute better option. When evaluating, enterprises can reference three metrics: whether the loop is locked to a single model vendor, how far the automation extends across the development pipeline (pure generation vs. connecting through to review and deployment), and whether the team itself has the capacity to take on the operational responsibility that comes with self-hosting an open-source model. The right answer depends on an enterprise's own technical capacity and governance needs, not on one route being inherently superior.

Source

The material that triggered this topic is a keynote compilation video published by the YouTube channel AI Engineer, “WF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF” (published around July 3, 2026, with view counts in the roughly 69,000–80,000 range across cross-verified third-party sources, not an official precise figure). This piece does not use that video's transcript as its factual basis; instead, it draws on Latent Space's on-the-ground reporting of AI Engineer World's Fair 2026 (June 29–July 2, 2026, Moscone West, San Francisco) — the AIEWF Daily Dispatch series — cross-checked against other media coverage, reorganized here as a point-by-point comparison rather than a sentence-by-sentence translation of the original reporting or the video. All attributions of statements to Microsoft, OpenAI, Z.ai, MiniMax, and Hugging Face representatives in this piece are limited to what that verified reporting covers; any specific technical specs, performance figures, or full quotes not provided in that reporting are not fabricated or extended upon here.