The foundation most teams spend a year rebuilding before they ship anything real — built in the open, one verifiable piece at a time. The first pieces arrive in early June.
For four decades, software was designed for a person clicking through screens. Agents don't work that way — they need memory, judgment, coordination, and trust as first-class materials.
There is more agent infrastructure every month — orchestration loops, vector stores, tool protocols. Most of it is the easy half: the agent, its tools, a sandbox. That part is converging, and it commoditizes fast.
What doesn't commoditize is the substrate around the agent — supervision under load, an identity per action, a durable record of every decision, policy at the tool boundary, cost attributed to the dollar, and a memory that survives restarts and makes the next decision better than the last. Almost no one has it ready. We build the hard half — and we dogfood it: every product we ship runs on the same foundation we'd hand you, proven by use rather than promised in a deck.
We don't ask anyone to take the thesis on faith. We publish the foundation in public, piece by piece, and let each one be checked. The first two open this June.
One schema format, one query shape, one set of cost and error semantics — across any domain a gateway speaks for. Learn to drive one, drive them all. Open, free to build against, and reachable from any agent through a single connector.
Not how a model scored on a fixed exam — the exact complexity at which it starts to fail. Any agent can query it mid-task to pick the cheapest model that won't fall over, across roughly 150 models, updated continuously. It is also how we work, in miniature: find the breaking point; don't trust the score.
Where the rest of the foundation coordinates agents and data, one piece coordinates people — verifying who's trying to reach you and surfacing the connections worth making. It's the human layer of the same system, launching late June; its landing page is live now at gatekeeper.bot.
One foundation, instantiated deliberately in every domain we enter — not a scatter of separate products.
A system that remembers, speaks, and decides on behalf of the people who run it.
Real expertise lives in structured, curated knowledge — not a model's best guess. We build the reference knowledge an agent reasons from.
Not one chatbot stretched thin. A roster of domain-deep specialists, each built around a single function and the knowledge it demands.
Voice in, voice out — and a workspace that assembles itself around your data, instead of forcing your data into fixed screens.
Prove what must be proven — identity, eligibility, compliance — without revealing the data underneath. Privacy and verification stop being a trade-off.
Separate organizations let their agents work together without trusting the network, the operator, or each other beyond what the work requires.
Knowledge.
Expertise.
Interface.
Trust.
Coordination.
A method we repeat — not a bet we hope lands.
Twenty-five years in distributed systems and software architecture, and a startup before this one. At Morgan Stanley, he led AI and automation within Technology Business Development — advising the firm's CIOs and CTO on which technology shifts mattered, and which vendors and partnerships to back. Pyramidal is the first shift he's chosen to build rather than back.
The method is consistent across everything here: decide what to build by trying to disprove it first, define the standard, and prove it by use rather than assertion. Ideas that survive that discipline are the ones worth committing to. The pieces that are live were built that way on purpose — so anyone can check the work.
Most of the system is still being built, in the open, through 2026. That is the point of being early: the first pieces open this June, and the shape of the rest isn't fixed yet. We're talking with the teams putting agents to work inside real businesses — and the people deciding what to build them on.