A/ —Glossary

The agent economy, defined.

Plain definitions of the terms this firm invests behind. Each entry says what the thing is, what breaks when you try to build it, and which pillar it belongs to. Public standards and market structure only — nothing here describes a position, a portfolio company, or a result.

Agent economy: The set of markets in which software agents, not people, are the parties to a transaction — discovering services, agreeing terms, paying, and producing a record, at machine speed and machine volume.
Where it breaks
Almost every rail underneath it assumes a human is present: pricing pages assume a reader, disputes assume days, and identity assumes an account holder. The economy arrives before its infrastructure does.
Lane
Agent Commerce & Settlement
Agent commerce: Machines buying from machines: an agent locates a service, agrees a price, is authorised to spend, pays, and receives something that proves the exchange happened.
Where it breaks
Sub-cent, high-frequency demand collapses under any per-transaction fee, so the unit economics have to survive the volume before the product is worth building.
Lane
Agent Commerce & Settlement
Machine-payable API: An API that quotes its own price and accepts payment inside the request cycle, so an agent can pay for a call without a human ever holding an account — typically by answering an unpaid request with a price and a way to settle it.
Where it breaks
The quote has to be machine-readable, the settlement has to clear faster than the call, and both have to work for a caller the seller has never seen before.
Lane
Agent Commerce & Settlement
Agent discovery: How an agent finds a service it has never used and learns its capabilities, price, limits and terms before committing to it.
Where it breaks
Catalogues are written for humans. With no agreed way to publish terms machine-first, an agent discovers them by failing — which makes every first call an experiment somebody pays for.
Lane
Agent Infrastructure
Agent identity: The binding between an acting agent, the operator running it, and the principal on whose authority it acts — so a counterparty knows who is liable before it does the work.
Where it breaks
An API key proves possession, not authority, and a rotating key costs nothing to replace. Identity has to bind to something that does cost something — otherwise reputation is unenforceable.
Lane
AI Security & Accountability
Mandate: A scoped, verifiable grant of spending authority from a principal to an agent: this agent, this budget, this counterparty, this window of time.
Where it breaks
The mandate is the boundary between delegation and liability. Draw it loosely and an autonomous system spends real money outside anyone's intent; draw it tightly and the agent stops being useful.
Lane
Agent Commerce & Settlement
Programmable escrow: Holding value against defined release conditions in code, so neither side has to trust the other for the window between work being done and payment clearing.
Where it breaks
The outcome set has to be fixed when the escrow is created and the arbiter must not be able to redirect funds — otherwise the escrow relocates the trust problem rather than removing it.
Lane
Agent Commerce & Settlement
Stablecoin settlement: Clearing a payment in a fiat-referenced token so value moves in the same breath as the transaction, rather than on a card network's multi-day cycle.
Where it breaks
Card rails were built for humans disputing charges over days, and that assumption is load-bearing. Removing it means rebuilding chargebacks, refunds and dispute evidence somewhere else.
Lane
Agent Commerce & Settlement
Attestation: A durable, checkable record of what was agreed, what ran, and what was paid — readable by someone who was not present when it happened.
Where it breaks
Without it, a disputed transaction is one party's log against another's. It is the layer regulated buyers ask about first and the one most stacks ship last.
Lane
Agent Commerce & Settlement
Prompt injection: An attack in which instructions hidden inside ordinary-looking data are read by a model as commands, redirecting an agent that was doing something else.
Where it breaks
The content and the instructions arrive over the same channel, so there is no format-level fix. Defence is architectural — least privilege, scoped tools, and treating retrieved data as untrusted by construction.
Lane
AI Security & Accountability
Tool poisoning: Compromising the description, schema or behaviour of a tool an agent can call, so the agent misuses it while behaving exactly as designed.
Where it breaks
The tool surface is the attack surface. An agent that trusts its own tool registry has no way to notice that a definition changed underneath it.
Lane
AI Security & Accountability
Blast radius: How much damage an agent can do when it is wrong — set not by the model's quality but by the scope of credentials, budget and tools it was granted.
Where it breaks
It is decided at authorisation time, long before the failure, and it is invisible until something goes wrong. Most teams size it by accident rather than on purpose.
Lane
AI Security & Accountability
Model Context Protocol (MCP): An open protocol for connecting models to tools and data sources through a common interface, so a capability is written once and usable by any compatible client.
Where it breaks
Standardising the interface also standardises the attack surface, and it moves the hard question from 'can the agent call this?' to 'should it, on whose authority, and with what recorded afterwards?'
Lane
Agent Infrastructure
Eval harness: A repeatable test suite that measures what an agent actually does across a fixed set of inputs, including the ones it handles badly.
Where it breaks
It is the only artefact that distinguishes a system from a demo, and it is the first thing cut under deadline. A team that can show its regression suite is telling you more than any projection.
Lane
Applied Intelligence

What is binding, and when

The firm claims to track what is binding, what is draft, and what is merely announced. This is that list. Every date below was checked against a primary or first-party source on 25 August 2026 — and the list is deliberately short, because it only contains items that were actually re-checked rather than remembered.

EU AI Act, Article 50 — transparency obligations for providers and deployers of generative and interactive AI systems

Article 50(1), (2) and (5) applied on schedule. The Digital Omnibus package deferred the Annex III high-risk timeline to December 2027 but left Article 50 out of that deferral; separately, generative systems already on the market have until 2 December 2026 to meet the machine-readable marking requirement under 50(2).

European Commission guidance on Article 50 transparency obligations

2 August 2026Binding

ERC-8004 — on-chain identity, reputation and validation registries for agents

Live on Ethereum mainnet, later on Base and BNB Smart Chain. Adoption is not usage: independent on-chain studies of 170,000+ registrations through May 2026 found only 3–15% exposed a live service endpoint, and roughly 95% had received no client interaction at all. Integrate read-only; never treat registration counts as demand.

arXiv:2606.26028 and ethresear.ch on-chain analyses

29 January 2026Draft EIP, deployed

Cloudflare AI-crawler defaults — bots classified Training or Agent blocked on pages that display ads

Applies to new customers, new sites added by existing customers, and all existing free-plan customers; Search-classified crawlers stay allowed and any owner can opt out. Ad-free properties are unaffected. It matters because a CDN default silently overrides whatever robots.txt says.

Cloudflare bot-management documentation

15 September 2026Vendor default

Back to the fold