BAML: New Programming Language for AI Era
Vaibhav Gupta, founder of YC-backed BAML, discusses building a programming language optimized for LLMs and AI agents. The discussion covers unifying type systems across code and data, shipping code at agent speed, and the strategic shift from SaaS to PaaS models.
The software industry stands at an inflection point where AI agents can generate code at machine speed, yet human verification remains the primary bottleneck. Vaibhav Gupta, founder of YC-backed BAML, argues that the solution lies in redefining how we structure, verify, and ship code through specialized infrastructure designed for the non-deterministic era.
BAML: The Type System for the AI Era
BAML addresses the fragmentation between codebases and data layers by introducing a shared type system that spans frontend, backend, and database partitions. Unlike legacy solutions like Protobuf, which suffer from poor developer experience, BAML offers an embedded language compatible with Python, TypeScript, Go, and Rust. This architecture enables seamless model swapping, rigorous testing of LLM functions, and automatic versioning of data structures. By treating types as first-class citizens across the entire stack, BAML eliminates the 'data trench' that plagues modern AI applications, ensuring that schema changes propagate safely without polluting historical data. The language is open source, while the cloud system provides observability and cost-reduction algorithms that optimize latency and expenses across both open and closed models.
Redefining CI/CD for Autonomous Agents
Traditional Continuous Integration/Continuous Deployment pipelines are ill-suited for agent-driven development, where code changes occur at volumes humans cannot review. Gupta proposes a new paradigm centered on granular trust and visualization. Instead of binary merge decisions, teams should implement 'time-dimension' locks that restrict changes to specific API signatures or critical modules while allowing agents to auto-merge peripheral updates. Furthermore, code reviews must evolve from line-by-line text scanning to graph-based visualizations of control flow. These visual diffs allow engineers to assess structural changes rapidly, immune to superficial refactoring, thereby enabling shipping velocity that matches agent generation speed.
Strategic Shifts: PaaS, Design, and Infinite Demand
The economic model of software is transforming alongside technical workflows. Gupta predicts a long-term migration from rigid SaaS products to flexible Platform-as-a-Service (PaaS) architectures. Future platforms will succeed by allowing customers to inject custom data models and logic into the orchestration loop, effectively turning the platform into a harness for user-defined workflows. This shift necessitates usage-based pricing models and emphasizes design as a critical differentiator. As AI lowers the barrier to entry, the number of software builders will expand exponentially, requiring interfaces that cater to non-technical users through interactive, visual experiences rather than terminal-based tools. Organizations must prioritize tooling that enforces type safety and visualization to harness AI's productivity gains while maintaining system reliability.
Key insights
-
AI automates 90% of software engineering 'plumbing,' shifting the engineer's role toward high-level systems design and problem-solving. This transition increases individual developer velocity significantly, with metrics showing consistent quarterly growth in code output.
Impact: Companies can reduce headcount costs for routine tasks while accelerating feature delivery by reallocating engineering focus to architecture and complex logic.
-
BAML solves the 'data trench' by unifying type systems across code and data layers, enabling safe versioning and model swapping for LLM outputs. This ensures deterministic behavior from non-deterministic systems and prevents data pollution during schema evolution.
Impact: Reduces debugging time and infrastructure costs by enforcing type safety at the language level, while improving reliability for enterprise AI applications.
-
Traditional CI/CD cannot handle agent-generated code volume; shipping at agent speed requires granular code locks and visual graph diffs. Trust must be managed through selective human review of critical boundaries rather than full code scans.
Impact: Enables organizations to leverage autonomous coding agents without compromising security, drastically reducing merge times and deployment cycles.
-
The market is shifting from SaaS to PaaS, where platforms allow customers to inject custom data models and logic. This creates a harness model where the platform orchestrates user-defined workflows, driving usage-based pricing.
Impact: SaaS companies that evolve to PaaS can increase retention and revenue by offering deeper customization and integrating seamlessly into customer ecosystems.
-
As AI expands the builder base by 10x, design becomes more critical. Non-technical users will demand interactive, visual interfaces over terminal tools, making UX a primary competitive advantage.
Impact: Investing in intuitive design and visual abstractions will capture a larger market of non-engineer builders, driving adoption and reducing support overhead.
Action items
-
Audit current LLM integration testing and implement structured evaluation frameworks. Ensure all AI functions have deterministic output validation and automated evals to catch hallucinations and schema drift.
Impact: Improves reliability of AI features and reduces production incidents caused by non-deterministic model outputs.
-
Adopt type-safe AI layers that bridge code and data. Use tools that enforce shared type systems across frontend, backend, and database to prevent data pollution and simplify model swapping.
Impact: Reduces technical debt and maintenance costs while enabling safer experimentation with different LLM providers.
-
Redefine code review processes for agent-generated code. Implement granular merge locks on critical APIs and explore visual graph diff tools to review structural changes efficiently.
Impact: Accelerates shipping velocity by removing human bottlenecks in routine merges while maintaining control over sensitive system components.
-
Evaluate potential to shift from SaaS to PaaS. Assess if your product can allow customers to inject custom logic or data models, moving toward a usage-based pricing model.
Impact: Positions the business for long-term growth by offering deeper customization and aligning revenue with customer value creation.
Quotes
“90% of software engineering, well, in most jobs, 100% of software engineering is plumbing. If you're lucky to do any sort of thing beyond that, like systems design or assembly or any of these harder problems, even then, only 10% of your job is that. 90% of your job is still plumbing. And AI just takes that 90% and just makes it go away.”
“We can now generate code at machine speed. That is a solved problem. There's no human that can write code faster than a machine can today. It's impossible. But we still cannot ship code at machine speed.”
“I think what we need is like PaaS will work more likely. People want data as a service or platform as a service and nothing else... I want 90% of your system and 10% of it built exactly for my use case.”