Four Predictions For How AI Will Change Software in 2026

Every grew from $65-70K MRR to $150K MRR in 2025 after five years of flat growth. The inflection point came in June with the launch of Quora (AI email assistant), Dan's appearance on Lenny's podcast, and the first Cloud Code Camp. The company scaled from 5 to nearly 20 people while maintaining a consulting business alongside subscription revenue.

• •Revenue doubled from ~$65K to $150K MRR starting June 2025, growing 20%+ monthly
• •Key catalysts: Quora launch, Lenny's podcast appearance (22.63% growth that month), Cloud Code Camp
• •Built a 'meaty subscription' with multiple AI products (Quora, Sparkle, Spyro) under one price
• •Scaled from 5 to 20 employees while maintaining genuine culture and consulting business
• •Brandon's arrival in May 2024 marked the beginning of consistent 2-5% monthly growth

Every is transitioning from bottoms-up product development to a cohesive platform strategy. They're building unified infrastructure (single sign-on, billing) and planning a shared memory/context layer across all apps. The challenge will be maintaining genuine motivation while implementing strategic personas and product planning.

• •Hired Willie to build unified platform with single sign-on and billing across all apps
• •2026 focus: unified memory and context layer for personalized experiences across products
• •Challenge: balancing genuine interest-driven building with strategic product planning
• •Every's strength is building tools they actually need and writing about the process
• •Must avoid losing authentic motivation while scaling to strategic personas

Dan predicts agent-native architectures will become standard, with three levels: (1) anything users can do, agents can do, (2) anything the code can do, agents can do, (3) anything developers can do, agents can do. This enables in-app customization and bug fixes triggered by users, with Anthropic leading and Notion following this paradigm.

• •Level 1: Agent can perform any user action (Atlas browser demonstrates this with settings)
• •Level 2: Agent can trigger any code function, even those users can't access directly
• •Level 3: Users can request features/fixes and agents submit PRs or create custom versions
• •Every is already implementing this - building Discord bot that writes code and submits PRs
• •Anthropic (Claude Code) and Notion are leading; OpenAI still focused on traditional engineers

Designers with taste and vision who were previously blocked by needing developers will become powerful builders through AI coding tools. Lucas (Every's creative director) exemplifies this, 'vibe coding' apps to enhance his work. The challenge is making tools accessible beyond just the best designers who aren't intimidated by code.

• •Designers have been held back from building full experiences due to inability to code
• •Lucas (Every's creative director) now builds functional apps to improve his creative work
• •Cheap code + design taste = highly evocative, well-crafted experiences
• •Cursor adding visual front-end updates, but many designers still fear touching code
• •Need to abstract away code complexity to expand market beyond elite designers

Dan identifies three types of AI-era engineers: traditional engineers using AI as accelerant, vibe coders who don't understand code, and 'compound/agentic engineers' who've reinvented software engineering by delegating all coding to agents and managing them from a higher level. This third path is a distinct skill requiring trade-offs traditional engineers resist.

• •Traditional engineers using AI still read/write code and stay deeply involved
• •Compound engineers never look at code, fully delegate to agents, manage from meta-level
• •This requires sacrificing the craft of coding (which many engineers love) for new skills
• •Enterprise software serves traditional engineers; small companies pioneering compound approach
• •Traditional engineering jobs will persist long-term but slowly decline like newspapers vs internet

Brandon warns that deepfakes and AI-generated content will cause serious problems during the 2026 midterms. Social media companies heavily invested in AI aren't labeling AI-generated content, and average users (like Brandon's mom) can't distinguish fake from real. He predicts a major incident will force mandatory 'made with AI' labels.

• •Real example: Brandon's mom believed AI-generated gorilla video was real, refused correction
• •Social media companies not labeling AI content despite being invested in AI success
• •Deepfakes (word rarely used now) will likely impact 2026 midterms or 2028 presidential election
• •Prediction: Grotesque fake videos of politicians will trigger mandatory AI labeling requirements
• •Disney-OpenAI deal training users to expect non-real content, normalizing AI-generated media

Discussion of when AI becomes native to non-technical workers' daily tools. Companies will blame layoffs on AI in 2026 (mostly as excuse), but real adoption requires AI integration into workplace tools like Excel. AGI timelines have shifted - while 2025 was successfully the 'year of agents,' true AGI (models running indefinitely with autonomy) remains distant and poorly defined.

• •Non-technical workers (example: Twilio deal desk) still don't find AI useful in daily work
• •AI needs to be embedded in workplace tools (Excel, etc.) for mass adoption, not separate ChatGPT
• •2026 will see AI-blamed layoffs that are actually about underlying business issues
• •Sam Altman's 'less than 1000 days to AGI' timeline has shifted; progress is fast but AGI is harder than expected
• •AGI redefined: models that run indefinitely without turning off, not just smart chatbots

True AI autonomy requires models to run indefinitely like children growing more independent over time. Current models can run 20-60 minutes but diverge from human common sense. Alignment training (making models predictable) conflicts with autonomy - models need agency and room to fail. 2026 will see new training approaches allowing more autonomous exploration.

• •AI autonomy follows childhood development curve: 5 minutes → 20-60 minutes → indefinite
• •Current limitation: models develop quirks that diverge from human decision-making over time
• •Alignment training (making models predictable) is essentially 'telling you what you want to hear'
• •True autonomy requires models having individual agency and permission to make mistakes
• •Prediction: 2026 will introduce training approaches and architectures enabling autonomous exploration