The foundation: engage.re
Shared Meaning 30 July 2026 11 min read

AI Agents Need a Shared Language. A UK Company Has Built One.

Four protocols now compete to connect AI agents, and NIST has opened an initiative to align them. Each one moves messages between systems, and none of them moves meaning. The industry has attempted the meaning problem for twenty-five years, and the attempts share one defect. Sense Future built a platform that holds meaning as data and enforces it at the server. This article explains the difference, gives the measurements, and states what your own system inherits.

Connect with us about connected systems · · Replies within 24 hours

4
competing agent protocols, none carrying business meaning
+17 to +23
points of model accuracy from supplying meaning
Dec 2025
engage.re in production, with a sovereign contract live since 6 July 2026

What the protocols solve

Four protocols compete for adoption. MCP connects a model to a tool. A2A lets one agent discover and call another. ACP and ANP cover agent messaging and network discovery. A published survey of agent interoperability protocols compares all four.

NIST announced its AI Agent Standards Initiative in February 2026, to work towards interoperable and secure agents. Research examines what the protocols leave open, in governance gaps in agent interoperability protocols.

Each protocol solves transport, discovery and the call contract, and each one solves those well. None of them says what a field means.

What a protocol does not carry

Consider two systems that both hold a customer. System A calls it a client, with a status of live, pending or closed. System B calls it an account, with a state of 1, 2 or 3.

A protocol moves both records faithfully. Somebody must still write the rule that says live equals 1, and that rule sits in code between the two systems.

The count of such rules grows with the square of your systems. Ten systems need up to 45 mappings, and each one holds a business decision that a developer made once.

A wrong mapping produces a valid record. No error appears, so the discovery comes months later from a person who notices a strange figure. Nobody can then say which records the error touched.

We give the protocol detail in MCP and A2A move messages, do they carry meaning.

The twenty-five-year attempt

The industry has worked on machine-readable meaning since the late 1990s. The World Wide Web Consortium published RDF for facts, OWL for concepts and rules, and SKOS for term hierarchies. All three work as specified.

Adoption stayed narrow, and one property explains why. Every one of those approaches describes data from beside the data.

A description held beside a system can fall out of step with the system, and nothing breaks when it does. The reports keep running, the dashboards keep loading, and the numbers slowly stop meaning what the document says. Nobody notices for a year.

Recording meaning has therefore always been optional in practice, whatever a project intended at the start. We compare the four approaches in semantic layer, ontology or knowledge graph.

What changes when the server enforces it

Sense Future built engage.re on a different arrangement. Meaning is not a document beside the data, and it is a condition of storing the data.

An application declares its vocabulary before it writes anything. Every record type, every field and every relationship. The server then refuses any write outside that declaration.

Four consequences follow, and each one is testable.

  • The description cannot drift. Data cannot exist in a form the dictionary does not describe, so no separate document needs maintenance.
  • Two systems agree without integration. Every concept holds a permanent identifier, generated deterministically, so two applications mean the same thing with no synchronisation between them.
  • A definition survives a rename. A label can change, and the identifier does not, so history stays intact.
  • A wrong meaning fails at the write. The failure appears at the moment of the mistake, rather than in a report months later.

The platform holds everything in five database tables, and those five never change. A new record type adds a dictionary entry instead of a table, so no migration follows a new requirement.

Has anybody measured the effect?

Yes, and independently of this platform. A 2026 benchmark gave three frontier models 99 business questions on a public dataset, twice each. The second run added a 4KB document describing the measures and the conventions.

Model Schema only Schema and meaning Change
Claude Opus 4.750.5%67.7%+17.2 points
Claude Sonnet 4.646.5%68.7%+22.2 points
GPT-5.445.5%68.7%+23.2 points

With the document present, the three models became statistically indistinguishable from each other. Without it, they also became indistinguishable.

The conclusion in the researchers' words. "the most consequential architectural decision is not which frontier model to use but whether the system is grounded in authoritative business semantics at all." From Semantic Layers for Reliable LLM-Powered Data Analytics.

A 4KB document produced that gain. An enforced dictionary supplies the same information and cannot go out of date. We set out the whole evidence base in why AI gets your own business data wrong.

What an AI agent receives

An agent working on a system built this way receives five things it usually has to guess.

  1. The meaning of every field. The dictionary states it, and the server enforces it.
  2. A stable identity. The agent holds its own identity, tied to a named human supervisor.
  3. A narrow permission. One action on one scope, cancellable on the next call.
  4. A record of what it did. Every access is an event in a signed chain.
  5. Current figures. A change reaches a reported number within 60 seconds.

Points two and three matter to a regulator. Singapore's governance framework for agentic AI requires each agent to hold a unique identity tied to its supervisor, and we cover that in your AI agent needs its own identity.

What ESRE Media offers

We build your system on engage.re. ESRE Media is the consultancy arm of Sense Future, and we work in Enfield, London.

Your system therefore arrives with these properties on its first day.

  • Your vocabulary is declared and enforced. Your definitions live in your records rather than in a document somebody maintains.
  • Your systems agree without integration. Deterministic identifiers mean two of your applications share a concept with no mapping between them.
  • Your system does not become legacy. Five tables that never change, so a new requirement adds a dictionary entry rather than a migration.
  • Your agents are accountable. One identity model covers your people, your applications and your agents, with narrow grants and a signed event log.
  • You can leave. Six interfaces over JSON, a published conformance package at version 2.0.0, self-hosting, and your own encryption keys.

The platform has run in production since December 2025, and its sovereign contract has been live since 6 July 2026.

Between organisations, and with regulators

Two organisations that both build on the platform share their concepts by construction. A council and a care provider mean the same thing by a person, so a data-sharing agreement then covers permission rather than translation.

A regulator asking for evidence receives records that carry their own definitions. The answer to what a field means comes from the data, and not from a person who remembers the project.

Sense Future sets out the wider thinking in Machines of Human Grace.

The same foundation at two sizes

A county council runs housing, revenues, social care and waste as separate systems, and each one holds a person. Every pair needs a mapping, and the council maintains six. On a shared dictionary the four services hold one concept of a person, and no mapping exists to maintain.

A care home group runs a records system, a rota tool and a family portal. Three systems, three mappings, and the same arithmetic. The registered manager cannot get one occupancy figure today, and the reason is identical to the council's reason.

The care home software and charity software pages set out the sector detail. Our guides to charity data management and multi-academy trust systems show the same problem inside one organisation.

What we do not claim

  • The protocols are not a mistake. They solve transport, discovery and the call contract, and those problems needed solving.
  • An enforced dictionary does not remove integration with systems you do not control. It removes it between systems that use the dictionary.
  • Recorded meaning does not make AI correct. The benchmark reached 68.7% with the document present, and that figure still needs a person to check it.

What to do next

  1. Count your integration mappings. That number is what a protocol will not remove.
  2. Take one field two systems both hold, and compare its permitted values.
  3. Ask where the meaning of your data is written down. Where the answer is the code, you hold no record of it.
  4. Write one page defining your five most-used measures.
  5. Then decide whether that page stays a document, or becomes a declaration your server enforces.

Common questions

Do AI agent protocols solve the data meaning problem?

No. MCP, A2A, ACP and ANP each solve transport, discovery and the call contract. None of them says what a field means. One system may call a record a client with a status of live, and another an account with a state of 1, and somebody must still write the rule that connects them.

Why did the Semantic Web not solve this?

Because RDF, OWL and SKOS all describe data from beside the data. A description held beside a system can fall out of step with the system, and nothing breaks when it does. Reports keep running while the numbers stop meaning what the document says, so recording meaning stayed optional in practice.

What does an enforced dictionary change?

It makes the description a condition of storing data. An application declares every record type, field and relationship before it writes anything, and the server refuses a write outside that declaration. The description cannot drift, two systems agree without integration, and a wrong meaning fails at the write rather than in a report months later.

How much does shared meaning improve AI accuracy?

By 17 to 23 percentage points in a 2026 benchmark. Three frontier models scored 45.5% to 50.5% with the schema alone, and 67.7% to 68.7% with a 4KB document describing what the data means. With the document present the three models became statistically indistinguishable from each other.

Is engage.re in production?

Yes. Sense Future built the platform, it has run in production since December 2025, and its sovereign contract has been live since 6 July 2026. A conformance package at version 2.0.0 is published so that a third party can build a conforming implementation.

What does ESRE Media build?

We build your system on engage.re. Your vocabulary is declared and enforced, your applications share concepts without mapping between them, five tables that never change mean no migration follows a new requirement, your agents are accountable under one identity model, and a published conformance package with your own keys means you can leave.

Connect with us about connected systems · · Replies within 24 hours

Sources and further reading