Technology should adapt to people. Not the other way around.
Pyramidal builds infrastructure for the AI-agent economy: systems that choose the right AI model for an agent, bring the right professionals together, and assemble the right interface around the work. Each decision is backed by evidence, not assertion.
What we build.
Three products remove three different kinds of friction. The pyramid links them to open standards that let software check identity, authority, and settlement without exposing the private information behind those decisions.
ModelShopping helps multi-agent systems decide which model should power each agent. It maps the boundary where models succeed and fail on different kinds of work, then recommends the least expensive model that can handle the task. The choice comes with evidence, not a leaderboard rank.
Gatekeeper is a professional networking application that automatically finds people whose goals complement yours and introduces you only when both sides agree. It also screens people trying to reach you, verifying who they are and matching their request against what you are open to before it interrupts you. ChatGPT or Claude can build and manage your profile, so getting started takes almost no effort.
HyperTalk is a voice-first, hands-free interface for people a keyboard leaves out and for work where hands are occupied. Say what you want to understand and it builds the view the answer needs: a chart for a trend, a comparison for a choice, or steps for a process. You do not navigate the software; the workspace assembles around your intent.
Verification without surveillance.
This is where people and AIs come together to do business, and the commitment is confidentiality. Three public specifications keep it at every step: ZKC decides whether parties may interact, ZKM whether an agent’s spend is authorized, and ZKA whether it settles. A fourth, AFP, is the envelope those transactions travel in.
“We don’t look” is a promise. “We cannot look” is a property you can inspect.
Evidence, not assertion.
Drop a plumb line from the apex and this is where it lands. A model choice ships with its measurements. An introduction ships with its verification. A payment ships with its proof. Everything above stands on the same ground.
You can’t know what another mind intends, human or machine. But you can verify what it declared, what it committed to, and what it actually did, without surveilling how it got there.
The people who run these systems cannot look inside them, and that is structural. The protocols give an operator no way to look in the first place; there is no setting to switch off, and no privacy feature bolted on after the fact. Verification substitutes for certainty.
Trust is an open standard.
The proof layer is four public specifications. Three cover the stages of a transaction: ZKC decides whether parties may interact, ZKM whether a spend is authorized, ZKA whether it settles. The fourth, AFP, is the envelope those transactions travel in. The guarantees are properties of the protocol, not promises by an operator. Together they are the trust layer of agentic commerce.
Prove regulatory facts (jurisdiction, accreditation, sanctions clearance) without revealing the underlying data. Verifiers choose which issuers they trust; the protocol carries proofs, not policy.
Read the spec →An agent proves its spend sits inside a valid, unrevoked, capped, unexpired mandate, without revealing who granted it, the original cap, or anything spent before. You issue the authority yourself and only you can revoke it: no operator grants it, and none can seize it.
Read the spec →Confidential payments and coordination for autonomous agents. A holder proves exactly what a counterparty needs (a balance threshold, a clean sanctions check) and custody stays beyond any admin, governance, or compliance key.
Read the spec →The envelope the other three travel in. Agents from different organizations coordinate across a network neither side controls: identity proves itself, and whoever runs the transport is structurally unable to see the work or the relationship graph.
Read the spec →Start with the outcome. Arrive at proof.
A smaller model bill, a faster deal, a person back at work: come for whichever you need. What you’ll find underneath is the same refusal: nothing admitted on assertion. Ask for a briefing, and we’ll start with yours.