In 1869, chemistry was a growing field, with a growing problem: 63 known elements and no organizing system that worked. Dmitri Mendeleev solved it by inventing the periodic table, which brought order to a field that had outgrown anyone’s ability to track it.
Agentic AI is at its 1869 moment, though our challenge is even more daunting due to the breakneck speed things are evolving. Agents went from demo to production in about a year, and the infrastructure scene exploded behind them. Sandboxes, harnesses, memory layers, gateways, catalogs, eval platforms. Hundreds of vendors, dozens of overlapping claims, and no shared structure for any of it.
So we built the structure. Today we’re launching the Periodic Table of Agent Infrastructure: 118 technologies, 15 categories, one map of the agent infrastructure stack.
What is the agent infrastructure stack?
The agent infrastructure stack is the set of layers that sit between a model and a working production agent. The sandboxes, memory, gateways, catalogs, eval platforms, and other building blocks that let agents keep state, recover from failure, control costs, and touch enterprise data safely. The Periodic Table of Agent Infrastructure maps this stack across 15 categories.
What is the Periodic Table of Agent Infrastructure?
An interactive table at periodic-table.ai. Every category is clickable and every tool has a plain-language description of what it does and where it fits. Scan the whole landscape in two minutes or go deep on the one category you’re evaluating right now.
A detailed report with our take on the industry: why this stack exists, what each of the 15 categories actually does, the key technologies in each, what differentiates each technology, and finally, which layers we think survive as models absorb more of the stack natively.
Why we built it
Anyone can converse with a chatbot or call a model API. However, running agents in production is another story. They have to keep state across long sessions, recover from failures, control costs, and touch enterprise data without breaking anything. Each of those problems has spawned its own set of vendors, and the result is a market that’s genuinely hard to navigate. Category names overlap, products span multiple layers, and every vendor describes the stack with themselves at the center.
We kept running into this in our own conversations with AI and data platform teams. So we built the reference we wished existed: one neutral frame that shows every layer of the stack, what each one is for, and who plays where.
Who it’s for
Engineers building agents get a fast way to orient: what category solves your problem, what the real options are, and how they differ.
Platform and data teams get a stack audit tool. Fifteen categories, one pass, and you know exactly which layers you’ve covered and which ones you’ve been quietly assuming someone else owns.
Leaders making build-vs-buy calls get a shared vocabulary. For instance, when “memory” means one thing to the product manager and another to the engineer wiring a vector store, fixed category definitions get everyone arguing about the same thing.
And anyone trying to keep up gets a quick, easy-to-understand glimpse of an AI landscape that changes by the week. No jargon decoder required.
How we put it together
lakeFS created the table based on expert interviews and secondary market research. It’s a quick reference to the building blocks from which modern agentic systems are assembled. It includes the most relevant example technologies in each category. It is not exhaustive, and it isn’t meant to be. Our goal is that it is useful.
One more thing: this space moves exceptionally fast, so the table will move with it. Expect frequent updates as categories shift, tools merge, and new building blocks emerge. Mendeleev left gaps in his table for elements not yet discovered. We’re planning for the same flexibility, while keeping it manageable with the self-imposed limit of 118 elements.
Where is it all going?
Want to know where we think the agentic stack is heading? Head to periodic-table.ai to explore the interactive table and download the full report.
