SEAM

Help agents find what they need

SEAM stands for Semantic Entities, Associations and Metrics. It is our working philosophy for giving AI agents a small amount of shared knowledge so they can quickly find the right sources, understand what they mean and use the right tools.

See how it works

Start with a real question

A new task often starts with the same questions: where is the information, which source should we trust and how do we reach it? SEAM records the useful definitions and connections so each new workflow has less to rediscover.

We start with questions people need answered. We add structure where it helps an agent find and use the evidence.

Semantic Entities, Associations and Metrics

What SEAM describes

Semantic means meaning. These three parts help explain what we are looking at, how it connects and what we measure.

E / Entities

What are we talking about?

Name the things your business works with and define what counts as each one.

Customer · Order · Campaign

A / Associations

How are they connected?

Record the relationships, so the connection between two things can be followed and checked.

A customer places an order

M / Metrics

What are we measuring?

Write down the calculation, the source and what is included or excluded.

Revenue from completed paid orders

A small starting point for a new workflow

Enough context to find the rest.

An agent starts with a small amount of context: enough to find relevant definitions and follow their links. It gathers more from the original sources as the task develops.

The benefit is speed: a workflow we have never run before can find its way around the organisation, connect the information it needs and identify the tools that can access it.

  1. 01

    A question

    What does someone need to know or do?

  2. 02

    A little starting context

    Find the relevant definition and its links.

  3. 03

    Follow the sources

    Read the linked records and search the original systems.

  4. 04

    Choose sources and tools

    Work out which meaning, source and available tool fit the task.

  5. 05

    Do and check the work

    Use the evidence, check the result and look further when needed.

A finding can lead back to another source or definition. The agent keeps looking as needed. Access still depends on the permissions of the tools it uses.

Different definitions can be right for different questions.

An analytics team may have several definitions of a customer, conversion or revenue figure across its systems. SEAM records what each means, when to use it and where to find it. Finding a matching term is only the start; the agent also needs to know whether it fits the question.

Where that helps

  1. 01

    Answer a business question

    Context to discover
    What the metric means, where it comes from and how the records connect
    Live information
    Warehouse, CRM or operational system
    What this enables
    An answer based on the intended definition, with the source made clear

  2. 02

    Audit an implementation

    Context to discover
    What should exist, which system is authoritative and what needs checking
    Live information
    GA4, GTM, consent configuration or data models
    What this enables
    Checks against what the setup is meant to do

  3. 03

    Prepare a client briefing

    Context to discover
    Which client, projects, people and documents belong together
    Live information
    Jira, Drive, Slack, CRM and delivery systems
    What this enables
    A briefing based on the relevant records and recent updates

  4. 04

    Carry out a task

    Context to discover
    What needs doing, who owns it and which tools can help
    Live information
    The tools authorised for that particular task
    What this enables
    An action within the agreed scope and tool permissions

Keep the records small and useful. Live facts stay in their source systems. Definitions and links need an owner and regular review too: copying more information creates more to keep up to date. Add detail when it helps answer a real question.

What it actually looks like

A definition you can read and follow.

We keep our definitions in Open Knowledge Format (OKF): Markdown files with YAML frontmatter. The YAML at the top names the record, its owner and its source. The text below explains the meaning, with links to related records.

This example describes an active client. SOW means statement of work. OpsHub is our operations system; HubSpot is our CRM. The file explains where to look. The current client records stay in those systems.

Delivery, finance and sales use different meanings of “active”. An agent needs to choose the meaning that fits the question, or ask which one is intended.

metrics/active_client.mdShortened extract from our maintained OKF bundle
---
type: Metric
name: active_client
title: "Active Client"
description: "A client with at least one active SOW."
layer: judgement
canonical_source: ops_hub
fallback_sources: [hubspot]
status: stable
owner: "mark@measurelab.co.uk"
review_cadence: quarterly
---

# Active Client

A client with at least one active SOW in OpsHub.

## Source

- Canonical: [ops_hub](../connections/ops_hub.md)
- Fallback: [hubspot](../connections/hubspot.md)

## Contexts

| Context  | Definition |
|----------|------------|
| default  | Any client with an active SOW |
| finance  | Any client invoiced in the last 90 days |
| sales    | Recent engagement or an open deal |

Where it stands

Our working philosophy

We use SEAM inside Measurelab for much of our agentic work. It works well for us: a new workflow can get started with less time spent finding its way around our systems.

We are still learning. The test is practical: does it help an agent find the right sources faster, make fewer wrong connections and need less correction?

We are happy to explore those questions with clients and collaborators who want to learn alongside us.

Let's talk

Working on how agents find and use knowledge in your organisation? We'd be happy to compare notes.

Get in touch