Specs precise enough for agents
Writing a problem down precisely enough that an agent can execute it — then building the agent, the tools it uses, and the infrastructure underneath. The spec is the deliverable; the code follows from it.
About
Aaron Chartier is a technologist working on agent orchestration, context engineering, and GenAI infrastructure and security — the unglamorous discipline of getting AI agents to do real work reliably.
His path runs from helpdesk through IT operations and process engineering into AI leadership: sixteen years in technology, the last several leading applied-AI work in a large enterprise. The operational background is the edge — he understood how enterprise systems actually break long before he started writing specifications for agents that take instructions literally.
His throughline is specification precision: vague intent is what makes agents fail in production, and writing a spec an agent can execute reliably is a learnable, teachable discipline. He pair-codes with agents — he specifies, the agent writes, he reviews for correctness and safety — evaluates and integrates the surrounding stack (Langfuse, LiteLLM, n8n, Azure OpenAI), and mentors developers from their first prompt to production agents.
He builds tools designed to be driven by agents — inkjet, a set of agent-first command-line tools, and this site, which is built to be read by machines as well as people.
Helpdesk → IT operations → process engineering → AI technology leadership — an operations-to-AI path, not a computer-science one. Sixteen years in technology, the last several leading applied-AI work.
Focus
specification precision agent orchestration context engineering prompt engineering LLM evaluation GenAI infrastructure AI security and guardrails token economics technical mentorship process engineering agent-first developer tooling CLI design
How I work
Legible, observable systems
Dull, legible things with clear seams and honest state — nothing clever where clear will do.
Ideas I work through
Specification precision is the bottleneck
Agents fail in production because the spec was vague, not because the model was weak. Writing specs an agent executes reliably is a learnable, teachable discipline.
The agent maturity progression
Assisted → captured → compressed playbooks → self-executing skills → autonomous. Not magic — a compression-and-pattern-recognition pipeline.
Agentic memory
Stop ten agents from independently making the same mistake ten times: agents that record learnings, compress them, and build on each other's.
The headless agentic web
Agents want fast structured data, not rendered pages — the web is being rebuilt for machine consumers. It's why this site ships a machine layer.
Process engineering for agent specs
Years spent documenting processes for humans to execute is the exact skill for documenting them for agents.
Stack I reach for
This site
TypeScript Astro Tailwind Vitest
Static output, fast.
Tools & ops
Python BLE Docker networking
From agent-first CLIs down to the containers and networking they run on.
Browser platform
Web Audio WebGL / GLSL WebGPU WebCrypto BigInt
The lab and tools lean on raw browser APIs — sound, shaders, an in-tab model, hashing, arbitrary precision.
Find me
Machine-readable version: about.md.