Insights
The reference · updated August 2026
A working reference for architects deciding whether a distributed ledger belongs in a system, and what it costs to run one inside an enterprise. Written from delivery experience, including the parts that went badly.
The complete reference: which decisions to automate, how to prove a system behaved correctly, what integration actually costs, and the failure modes that appear six months after a successful pilot.
v2.1 · 9 hubsA working reference for architects deciding whether a distributed ledger belongs in a system, and what it costs to run one inside an enterprise. Written from delivery experience, including the parts that went badly.
v3.0 · 8 hubsWhat an agent may commit, under whose approval, against which limit, and how the action is unwound. The reference for putting agents inside a control environment rather than beside one.
v1.2 · 7 hubsThe register, the restrictions, the redemption path and the reconciliation behind them. The engineering half of tokenizing an instrument, written for the team that has to operate it afterwards.
v1.4 · 8 hubsA working reference for the engineer asked whether agents can safely act across an organisation’s boundary. Nine questions an agent must answer to operate — who it is, what exists, who can act, whether to deal, how to reach the old web, how the organisation is shown, how it operates, how value moves, and what happened — each with the open spec that answers it and the failure it is written to refuse. Written by a practice that shipped through the token-sale era of 2017–2019 and now builds the protocols underneath it.
v1.0 · 9 hubsLicensing software someone else operates under their own brand: the responsibility split, the upgrade path, the support posture, and the evidence a licensing buyer checks before a demo.
v1.1 · 8 hubsAn engineering practice inside a merchant bank has an unusual reason to build open protocols rather than products: it consults on, builds, and licenses the same infrastructure, and the protocol is the part that is worth owning. What that means for how we work.
The last question an agent stack has to answer is what happened — in a record that outlives the vendor that produced it. Sealed receipts, verification computed rather than asserted, and a notary leaf that reveals a time and a kind and nothing about who.
An agent that can pay needs a policy that decides whether a payment goes, stops, or waits for a person. The commerce layer is a pure allow / deny / escalate evaluator over integer money — where a child policy may only narrow and a human gate on money is mandatory.
When two organisations’ agents finish a piece of work together, three layers close it: execution operates it, commerce settles it, and audit seals it. A reading of the estate’s open-ledger case study as the top of the agentic-internet stack.
How a regulated firm declares itself to machines: an AAO charter of roles, families and an accountable human, with approval gates only where a mistake would hurt. AAO expands to Agentic Autonomous Organization.
Most organisations an agent needs still live on the ordinary web. The gateway layer puts a machine door on an existing service: a discovery document at a well-known path, capabilities as verbs, prices in whole minor units, and endpoints that may not leave your own domain.
Deal flow is trust routing: who knows whom, and who will vouch. The trust layer of the agentic internet routes an introduction along consented relationships — and refuses, structurally, to let a decline reveal that a path ever existed.
A procurement team spends weeks finding who can supply a thing. The discovery layer lets an organisation publish what it wants next, one document at its own domain — with a visibility no caller can override and an expiry no caller can set.
Before an agent can transact with your organisation it has to find you and learn what you are. A worked path from a realm/1 registry entry to a graph/1 node — and why the graph refuses to carry a relationship you assert about yourself.
An enterprise agent needs authority it can prove and a parent that can revoke it. The identity layer of the agentic internet is a chain that runs human → organisation → agent → sub-agent → session and can never widen on the way down — because the spec refuses every widening by name.
The token-sale era of 2017–2019 gave software a way to hold and move value. It never gave an organisation an open, checkable answer to the five questions an agent acting on its behalf now forces: who is acting, on whose authority, whom to trust, may value move, and what happened.
A per-agent spending cap looks like a policy decision. It is also an integration question: does the SDK the agent’s wallet runs on enforce one at all? The answer is usually no, and finding that out costs a day if you look before you build, and a production incident if you don’t.
What changed in the architecture of mlgblockchain.com, and why the old service pages now point here.
A test suite built from labelled history, run before anyone sees a demo — and an honest account of what a passing score does not prove.
A ceremony performed once, by the people who designed it, proves very little. Designing one an operations team can run without you is the deliverable.
Who can change the code, under what approval, and what that means to an auditor.
A working example: idempotency keys, reorg handling, and a documented break procedure.
When an instrument is redeemable against a physical reserve, the custody boundary and the reconciliation model are one design.
How an accounting system’s finality assumption decides the shortlist before throughput does.
A 3-of-5 signing policy looks conservative on the day it is designed. It is a headcount bet, and most institutions lose it within eighteen months — quietly, because nothing breaks until the day a signature is needed.
The build vocabulary as a machine-readable term set. 103 terms published of a 250-term target, each defined once and linked from every page that uses it.
All termsllms.txt, plus TechArticle and DefinedTermSet structured data throughout. Nothing here is gated or paywalled, and citation blocks are stable.
Search the library