Stop coding, build Software Factories instead

Summary

David Ondrej argues that AI-assisted work is moving beyond chatting with agents or manually launching them one at a time. He proposes building software factories: coordinated sets of agent workflows that run repeatedly or in parallel, with less human intervention.

Main ideas

• Shift human effort toward higher-level work. Ondrej says agents are increasingly capable of coding, research, design, and routine business tasks. He argues that people should focus instead on judgment, taste, prioritization, practical constraints, and deciding what is worth building.
• People are the bottleneck in agent workflows. Manually prompting, approving, and starting every task limits how many agents can be used. He expects workflows to be activated by schedules, webhooks, product events, errors, customer feedback, and other signals—not just by a person typing a prompt.
• Automate repeatable work incrementally. Ondrej suggests examining recurring tasks, tracking where time goes, and automating a small, predictable portion of the work. He recommends expanding gradually and refining workflows when early runs produce poor or wasteful results.
• A software factory is a collection of agent loops. A loop has a trigger, an agent that performs a task using its model, skills, context, and tools, and some form of checking or follow-up. A factory combines multiple loops across areas such as product development, billing, database health, security, and customer support. Human approval may remain necessary for important decisions.
• Model choice and verification matter. Ondrej describes using different models for orchestration, implementation, and review—for example, assigning a stronger model to coordinate agents and cheaper models to routine subtasks. He also recommends having a model from a different provider review important work to catch issues or offer another perspective.
• Skills and prompts are building blocks, not the end state. Reusable prompt templates can speed up common requests, while “skills” encode more complete, repeatable procedures. Ondrej predicts that sequences of skills—potentially inferred from repeated patterns in agent sessions—could be a further step toward automation.

Local agents versus cloud agents

Ondrej argues that local agents are constrained by a user’s computer and can interfere with normal computer use: they share hardware, may open windows or tabs, and stop if the machine or internet connection is unavailable. By contrast, he presents cloud agents as running in isolated sandboxes that can be provisioned as needed, making large-scale parallel work and always-on triggers more practical.

He also argues that isolated cloud environments can limit the impact of a compromised agent compared with giving an agent access to a user’s main computer. He acknowledges that local execution can still be useful for tasks involving local files or privacy-sensitive work.

Cloudroom product demonstration

Ondrej presents Cloudroom, his platform for running agents locally or in cloud sandboxes. In the demo, he shows:

• Launching multiple subagents to inspect a codebase, each in a separate environment.
• Using a hierarchy of manager and worker agents, with different models assigned to different roles.
• Pausing, archiving, and restarting groups of agents with a different model or agent count.
• Creating an automation in natural language: every two hours, an agent checks the product against its architecture decision records, proposes an improvement, and sends the proposal to a separate model for adversarial review. The result is left for human review only if the agents do not agree to archive it.

Cloudroom is described as aiming to support multiple models, agent harnesses, subscriptions, and both local and cloud workflows. Ondrej also promotes Cloudroom Ultra, which he says includes personalized onboarding and 10,000 cloud-agent launches over 30 days. These are product and pricing claims made in the video.

Practical guidance and examples

The video’s main tutorial is a conceptual walkthrough of building a small agent factory:

1. Identify repetitive work and measure where time is spent.
2. Choose a clear, bounded task and an appropriate trigger.
3. Give the agent relevant context, skills, APIs, and tools.
4. Add checks or a second-model review where useful.
5. Test the automation, refine it, and expand cautiously.

Examples include periodic billing-code and database-health checks, automated triage of user-reported bugs, security reviews, codebase simplification, and identifying product improvements. Ondrej cautions against measuring success by generated files or code volume; the goal should be useful outcomes.

Analysis and claims

Ondrej predicts that agent fleets and cloud execution will become more common, and that the productivity gap between people who build effective agent systems and those who use only basic AI tools may widen. He compares automation’s effects to earlier technological shifts in agriculture and computing, while emphasizing that adapting work—not preserving every manual task—is the response he recommends.

He also makes comparative claims about the value of AI subscriptions and particular models, citing a SemiAnalysis study and his own usage. These are his assessments and may depend on the models, plans, and period discussed; the auto-generated subtitles contain several potentially inaccurate model names and figures.

Reviews, guides, and tutorials mentioned

• Guide: Start with a small, predictable task and automate it in stages.
• Tutorial/demo: Build a scheduled product-improvement automation in Cloudroom, then have a second agent review its proposal.
• Workflow examples: Billing checks, database monitoring, bug triage, security audits, and codebase reviews.
• Product demo: Run many isolated cloud subagents, coordinate them hierarchically, and manage them from one interface.
• Review approach: Have a separate model or provider evaluate important agent output.

Main speakers and sources

• Main speaker: David Ondrej, presenting his views, examples, and Cloudroom demonstrations.
• Sponsor discussed: PostHog, described as a product-analytics tool with session replay.
• External sources referenced by Ondrej: A Cursor study about cloud-agent usage, SemiAnalysis on subscription value, and a comment attributed to Vercel CEO Guillermo.
• Other materials mentioned: Ondrej’s open-source skills repository and a project called Director.