# Prospice Capital A venture partnership investing at the earliest stage in the infrastructure of the agent economy — the rails, identity, settlement and accountability that machine-to-machine work runs on. Toronto · San Francisco · Hong Kong. Source: https://prospice-capital.com/ --- ## 01 — Capital as foundation. Partnership as catalyst. We invest at the earliest stage in the agent economy — on a conviction the market still underwrites too cheaply: that machines transacting with machines will need rails, identity and accountability long before they need more intelligence. - **Stage**: Earliest - **Pillars**: 04 - **NAV core**: Agent Commerce - **Geography**: NA · Asia ## 02 — Investment Thesis: Where we deploy conviction. A focused allocation across the agent stack — weighted toward the layers that have to exist before autonomy is worth buying. | # | Pillar | % NAV | Detail | |---|---|---|---| | 01 | Agent Commerce & Settlement (Signature pillar) | 40 | Machines paying machines: machine-payable APIs, agent-readable discovery, programmable escrow, stablecoin settlement, and the receipts that make a disputed transaction resolvable. The least crowded layer of the stack, and our signature position. | | 02 | Agent Infrastructure | 30 | The substrate agents actually run on: durable runtimes and memory, tool protocols like MCP, orchestration and hand-off, and the observability that can trace one decision back through every hop that produced it. | | 03 | AI Security & Accountability | 20 | Prompt-injection and tool-poisoning defence, agent identity and scoped authority, and the audit trail a regulated buyer asks for on day one. Autonomy is sold on capability and bought on accountability. | | 04 | Applied Intelligence | 10 | Vertical agents that close a loop a human used to close — measured on the eval harness, not the demo. We look for teams who can show a regression suite rather than a highlight reel. | ## 03 — The Agent Transaction: One transaction, seven places it breaks. When software buys from software, the interesting problems are not in the model. They are in the seven hops between intent and proof. Select a hop. | # | What | Where it breaks | Lane | |---|---|---|---| | 01 Discover | An agent locates a service it has never used, reads its capabilities, and decides it is a candidate — without a human in the loop. | Catalogues are written for humans. There is no agreed way for an agent to learn price, limits and terms before committing, so it discovers them by failing. | Agent Infrastructure | | 02 Identify | Both sides establish who is acting, on whose authority, and with what scope — the agent, its operator, and the principal behind it. | An API key proves possession, not authority — and registration is not demand. Of the 170,000+ agents registered to ERC-8004 across Ethereum, BSC and Base by May 2026, independent on-chain studies found only 3–15% exposed a live service endpoint, and roughly 95% had never been used at all. A rotating key costs nothing; identity has to bind to something that does. | AI Security & Accountability | | 03 Negotiate | Price, constraints, latency and liability are agreed for this specific call, at machine speed and machine volume. | Human pricing pages assume a human reads them. Sub-cent, high-frequency demand collapses under any per-transaction fee — the unit economics have to survive the volume. | Agent Commerce & Settlement | | 04 Authorize | The principal's mandate is checked: this agent, this budget, this counterparty, this window. | Mandates are the boundary between delegation and liability. Get it wrong and an autonomous system spends real money outside anyone's intent. | Agent Commerce & Settlement | | 05 Execute | The work runs — tools invoked, context assembled, sub-agents dispatched, results returned. | The tool surface is the attack surface. Injection and tool-poisoning arrive as ordinary data, and blast radius is set by whatever scope you granted at hop 02. | AI Security & Accountability | | 06 Settle | Value moves — stablecoin rails, programmable escrow, or a fiat referral, at the speed the transaction happened. | Card rails were built for humans disputing charges over days. Machine cadence needs settlement that clears in the same breath as the call. | Agent Commerce & Settlement | | 07 Attest | A durable record of what was agreed, what ran, and what was paid — readable by someone who was not there. | Without it a disputed transaction is one party's log against another's. This is the layer regulated buyers ask about first, and the one most stacks ship last. | Agent Commerce & Settlement | ## 04 — Philosophy: How we actually work. ### 01 — We price the discount The market underwrites governable, auditable AI as a cost centre. We think it is the precondition for enterprise adoption, and we buy that mispricing at formation — before it is consensus. ### 02 — Evidence over narrative We would rather read your eval harness than your deck. A team that can show what its system does on the inputs it fails is telling us more than any projection. ### 03 — Few positions, held long Concentration is a form of respect. We would rather be the investor who answers on a Sunday for a handful of companies than a name on fifty cap tables. ### 04 — We read the code Threat models, tool permissions, audit trails, and the blast radius when an agent is wrong. Technical diligence done by someone who has shipped, not outsourced to a checklist. ### 05 — Security review, pre-round Injection surface, credential scope, data egress paths, audit trail. Findings go to the founder first — never used as negotiating leverage. ### 06 — Asia is a seat, not a slide Regional distribution, regulatory reading, and buyers who are genuinely reachable — plus an honest read on when not to raise at all. ## 05 — Team: Small on purpose. A concentrated partnership rather than a platform. The person who reads your repo is the person who stays with the company — there is no associate layer between a founder and a decision. Most of the work is not the investment itself but the shape of the table around it: who sits alongside you, what they are actually on the hook for, and how early that is set. We work from North America and Asia, which is a working constraint as much as a footprint — it means we are awake when your Asian customers are. - **Access**: Principals only - **Coverage**: NA · Asia timezones - **Horizon**: The whole arc - Toronto — North America — 43.65°N 79.38°W - San Francisco — North America — 37.77°N 122.42°W - Hong Kong — Asia — 22.32°N 114.17°E ## 06 — Contact: Send the repo. No deck required for a first conversation. A README, a design doc, or a demo that fails honestly tells us more than twenty slides. Tell us what you are building, what breaks at scale, and who you wish were in the room. You will get a real answer either way — including a no, with the reason. - frontier@prospice-capital.com - https://www.linkedin.com/company/prospice-capital --- No portfolio, AUM, or track-record figures are published on this site.