Software Factories

EY and 8090 just formalized what the industry has been circling for two years. Two parallel tracks are now reshaping how software gets built — and neither is waiting for the other.

What happened

On March 18, 2026, Ernst & Young LLP and 8090 launched EY.ai PDLC — an AI-native Product Development Lifecycle powered by 8090’s Software Factory. The announcement is significant not because EY built a new tool, but because of what it signals: a Big Four firm is replacing the traditional software development process with a mesh of AI agents operating across the full lifecycle.

The numbers from their internal use case: 70% increase in development productivity, 80x faster delivery, 95%+ automated test coverage. They plan to roll this out to tens of thousands of consultants at EY US.

That’s not a pilot. That’s an organizational commitment.

What a Software Factory actually is

8090’s Software Factory doesn’t generate code snippets. It orchestrates a “collaborative mesh” of AI agents with human oversight across the entire software lifecycle — requirements, architecture, implementation, testing, infrastructure, and operations. As Chamath Palihapitiya, cofounder of 8090, put it: “For 50 years, we’ve watched the same cycle repeat. A company initially writes their own software, then outsources it to a commercial vendor, then offshores the maintenance of that system, all while costs keep rising and quality keeps falling.”

The Software Factory model breaks that cycle by treating the entire development process as something an AI system can orchestrate end-to-end, not just the coding step. Unlike single-model AI tools, it coordinates across multiple AI models to ensure documentation, governance, and consistency at enterprise scale.

EY can deploy this because they already have the client relationships, the industry expertise, and the consulting infrastructure. 8090 brings the platform. Together, they address two primary use cases:

• Legacy modernization — systematically retiring technical debt, modernizing aging systems, reducing the operational burden of decades-old code
• New product development — building new software with the governance, quality, and consistency that AI-assisted development alone cannot deliver

The two parallel tracks

Here’s what makes this moment interesting: enterprise software factories and personal agentic tools are developing simultaneously, largely independent of each other. Neither is waiting for the other to finish.

Most enterprises today are still operating the way they always have: Jira boards, sprint cycles, handoffs between siloed teams, QA as a separate phase. That’s the incumbent model, and it still works — but it’s increasingly expensive and slow relative to what’s now possible.

The incumbents aren’t standing still out of ignorance. They’re standing still because changing how an entire organization builds software is genuinely hard. You have compliance requirements, existing vendor relationships, teams with established workflows, and institutional risk aversion. EY’s move is notable precisely because they’re willing to push through that inertia — and they have 8090’s platform to do it with.

Meanwhile, individual practitioners have already moved. Engineers and founders using personal agents, Claude Code, Codex CLI, or OpenClaw are operating at a level of autonomy and speed that most enterprise teams haven’t experienced yet. The gap isn’t closing — it’s widening.

Level 5/6 agentic engineering

The workforce question is the one nobody wants to talk about directly, but EY’s announcement makes it unavoidable. They’re planning to deploy EY.ai PDLC to tens of thousands of consultants. That means tens of thousands of people need to learn an entirely new way of working.

This isn’t “learn a new IDE” or “adopt a new framework.” It’s a fundamental shift in what the job is. At the highest levels of agentic engineering — what the industry is starting to call Level 5 and Level 6 — the human doesn’t write code at all. The human defines intent, sets constraints, reviews output, and orchestrates agents. The agent does the implementation.

• Level 5 — Full delegation with oversight: the agent handles the complete lifecycle from spec to deploy, the human reviews and course-corrects
• Level 6 — Multi-agent orchestration: multiple agents collaborate on different aspects of a system, coordinated by policy and human governance

EY’s Software Factory operates at Level 6. Their “collaborative mesh” of AI agents is exactly this — multiple agents working across requirements, architecture, code, test, and infrastructure, with human oversight at defined checkpoints.

But here’s the tension: the workforce doesn’t get there overnight. You can’t hand someone who’s been writing Java for fifteen years a multi-agent orchestration platform and expect them to be productive on day one. The mental model is different. The skills are different. The judgment required is different.

This is where the parallel innovation matters. The people who have already been working with personal agents — who’ve developed the instinct for agent delegation, context structuring, and output verification — have a head start. Whether they did that with OpenClaw, Claude Code, Codex CLI, or any other tool, the underlying competence transfers.

Comparing the approaches

What follows is a neutral comparison. Each approach has genuine strengths and genuine limitations. The right choice depends on context — organization size, risk tolerance, the problem being solved, and where the team is today.

OpenClaw is an open-source autonomous coding agent that gives the individual full control over the agent’s behavior, tool use, and model selection. It excels at end-to-end project delivery for individuals and small teams who want direct access to the raw capability without enterprise overhead.

Claude Code is Anthropic’s CLI-based coding agent. It’s tightly integrated with Claude’s capabilities — deep reasoning, large context windows, and strong code generation. It’s the most polished single-agent coding experience available today, purpose-built for engineers who want a high-quality agent they can trust with complex tasks.

Claude Co-Work extends Claude’s capabilities into collaborative, multi-step work. It moves beyond single-prompt interactions into sustained working sessions where the agent maintains context, manages tasks, and coordinates across workstreams. It’s Anthropic’s answer to the orchestration question — at the individual and team level rather than the enterprise level.

Codex CLI is OpenAI’s terminal-based coding agent — the direct counterpart to Claude Code. It’s the harness: you point it at a codebase, give it a task, and it executes. Same model, same workflow, different vendor.

OpenAI Codex App is the orchestration layer on top. It runs parallel threads with Git worktree isolation, built-in diff review, and a plugin system that extends the agent into external services — Gmail, Slack, Google Drive, and custom integrations via MCP servers. With automations, skills, and subagents, the Codex App is OpenAI’s play for the same orchestration space as Claude Co-Work: an agent that doesn’t just write code, but manages workflows end-to-end. Included with ChatGPT Plus, Pro, and Enterprise plans.

EY.ai PDLC / Software Factory is the enterprise-scale answer to the same question. It trades individual flexibility for organizational consistency, governance, and the ability to deploy across thousands of practitioners simultaneously.

The adoption question

EY’s advantage is distribution. They have the client relationships. They have the consulting infrastructure to drive organizational change. They have the credibility with boards and C-suites. When a Fortune 500 CTO needs to justify an AI-native development approach to their board, “we’re working with EY” is an answer that gets approved.

The personal tools’ advantage is speed and depth of skill development. An engineer who’s been working with Claude Code, Codex CLI, or OpenClaw for six months has already internalized the mental model that enterprise programs are still trying to teach. They know how to structure context. They know how to verify agent output. They know what breaks and why.

The question isn’t which approach wins. The question is how fast the enterprise track can close the skills gap — and whether the individuals who started early will be the ones leading that transition.

What’s clear is that the old model — manual development, linear handoffs, siloed teams — is now being challenged from both directions simultaneously. Top-down by firms like EY deploying platform-level automation. Bottom-up by individuals who’ve already adopted agentic workflows.

The companies caught in the middle — too large to move fast, too invested in existing processes to change easily — are the ones under the most pressure. Their engineers see what’s possible. Their competitors are adopting it. And now, their consulting partners are building it into the standard engagement model.

What this means for you

If you’re a Founder or CEO, the takeaway is structural: the way software gets built is splitting into two parallel tracks, and both are accelerating. The enterprise track will bring AI-native development to large organizations through platforms and consulting partnerships. The personal track will bring it to individuals and teams through open tools and direct agent access.

You don’t have to pick one. But you do have to start.

The skills your team needs — agent delegation, context engineering, output verification, multi-agent orchestration — are the same regardless of which platform they end up using. The earlier they develop those skills, the more options they’ll have when the enterprise platforms mature.