Building execution-grade AI systems
Agent architectures, durable workflows, and adversarial safety testing.
aaron@hertzfelt.io
AI Systems R&D · Founder, Hertzfelt Labs
Welcome. I'm an AI assistant with knowledge of Aaron's work.
Type a command or ask a question. Use /help for options.
Agent Architecture
Execution Flow
User Input
Natural language query
AI Gateway
Vercel AI Gateway
Reasoning
Claude / GPT-4 / Gemini
Tool Execution
Structured outputs + function calls
State Management
Durable execution context
Response
Streaming output
Vercel AI SDK + AI Gateway
Aaron Diemel
Founder, Hertzfelt Labs
I spent 15 years as an audio engineer—signal flow, systems thinking, shipping under pressure. In 2022, I made a bet: reskill entirely with AI as my copilot. No bootcamps, no traditional path. Just me and the models, learning to build together.
It worked. The mental models from audio engineering—signal chains, iterative refinement, translating vision into execution—transferred directly to AI systems architecture. I've scaled my capabilities at the same rate the technology has scaled, staying on the frontier while shipping production systems.
In 2023, I founded Hertzfelt Labs as a consulting practice and applied R&D lab. I work with founders and engineering teams on agent architectures, adversarial evaluation, and AI systems that actually execute. This site is built with the same tools and workflows I bring to client work.
Selected Work
Systems that execute, resume, and adapt.
How I Work With Agents
Speed comes from front-loaded architecture and domain translation, followed by delegation to coding agents. Here's the methodology behind shipping functional AI systems rapidly.
Architecture First
Front-load system design and domain translation before writing code. Define execution graphs, state schemas, and tool interfaces upfront.
Claude.md & Agents.md
Maintain project context files that give coding agents full system awareness. Persistent memory across sessions for consistent architectural decisions.
MCP Servers
Custom and OSS Model Context Protocol servers for tool integration. Extend agent capabilities with workspace tools, database access, and external APIs.
Agent Delegation
Delegate implementation to coding agents after architecture is set. Claude Code for complex refactoring, Cursor for iteration, v0 for UI prototyping.
Structured Outputs
Type-safe tool calling with Zod schemas and JSON Schema. Reliable structured extraction and function execution with validation at every step.
Evaluation & Iteration
Continuous evaluation against defined success criteria. Rapid iteration cycles with automated testing and human review gates.
Agent Tooling & Harness
# Project Context
## Architecture
- Durable agent execution with checkpoint-based state
- Tool calling via Vercel AI SDK with Zod schemas
- MCP servers for workspace and database integration
## Conventions
- All tools must return structured outputs
- Human approval gates for destructive operations
- State persisted to Supabase between sessions
## Current Focus
- Implementing document extraction pipeline
- Adding multi-model fallback for reliabilityStack
Tools and platforms in active use.
Languages & Frameworks
Agent Tooling
Model Inference
Development
Deployment & Infrastructure
ML Ops
Multimodal - Voice
Multimodal - Image
Recent Thoughts
Insights on AI systems, safety research, and frontier model deployment.
Everyone calling 2025 the year of agents clearly missed the memo. That was 2024. This year? It's about operational intelligence—knowing when not to deploy agents. True flex isn't scaling agent frameworks; it's understanding their limits. Sometimes a lean LLM stack with precision tool integrations will outperform a bloated agent architecture. 2025 is the pivot point—less about chasing agent hype, more about architecting for efficiency, modularity, and actual ROI. It's the year of strategic deployment, not blind adoption.
🚀 Pushing frontier #AI safety to the limit. Through rigorous adversarial testing, I stress-tested @OpenAI #o3 model before launch, exposing vulnerabilities that shaped its final safety architecture. Super stoked to see my work cited in OpenAI's system card, proving that #AISafety isn't theater—it's engineering.
