Introducing the Context Product

Manage your business context like a product.

A Context Product gives your AI agents the rules and definitions your business runs on. Each one has an owner and a version, and it passes its tests before it ships.

Get early access See an example

We are opening to a few teams first.

Scattered rules and definitions flow into one Context Product, Support triage, version 1.3.0, which serves three agents.
The problem

The answers exist. Your agents can't reach them.

Your business knows when a refund needs approval. But the answer lives in a prompt, a slide, a wiki page and someone's head. Each one says something different, and nobody owns it. So each agent picks a version, or invents one.

Support bot, system prompt

Team leads can approve refunds up to €15,000.

Finance review, slide 14

Any refund above €10,000 goes to Finance.

Support wiki, last edited in 2024

For large refunds, ask Finance first.

Written nowhere

Legacy contracts never need Finance.

One rule, four versions, no owner
57%

of enterprises in a July 2026 survey traced confidently wrong agent answers to missing or inconsistent business context.

VentureBeat VB Pulse, 101 respondents. Read the survey
1,000

context engineers: Cognizant is adding this many in 2026. The work already has a job title.

Cognizant, reported by Nasdaq. Read the article
An example

Same question. Two agents.

Both agents use the same model. The second one also reads a Context Product.

From a support ticket

A customer asks for a €12,000 refund. Can a support lead approve it alone?

Without a Context ProductMade up

Yes. Support leads can approve refunds up to €15,000.

No source and no owner. It still sounds sure.

With a Context ProductApproved rule

No. Refunds above €10,000 need a Finance approver. One exception: customers on a legacy contract.

Rule
Refund approval thresholds, version 4
Owner
Finance Ops
Source
Refund policy, section 3.2
Approved
2 October 2026
Freshness
Checked daily at 06:00 UTC

A fictional company. The figures are examples.

Definition

What is a Context Product?

A Context Product is a package of business context that one job can rely on. It has an owner and a version, and it answers its test questions before it ships.

  1. 1

    A job

    Each Context Product serves one job, such as triaging tickets or approving discounts.

  2. 2

    An owner

    A named person or team answers for it, and decides what it carries.

  3. 3

    Contents

    The definitions, rules, thresholds and policies the job needs. A rule selects them, so the product stays current.

  4. 4

    Tests

    Questions with verified answers. A new version ships only when it answers all of them correctly.

Context Product Support triage v1.3.0Published
1
Job
Route and prioritise support tickets
2
Owner
Support Ops
3
Contents
48 definitions, rules and thresholdsSelected by a rule: everything approved in Support
4
Tests
12 of 12 answered correctly
5
Version
1.3.0, published 6 OctoberChanged: the P1 rule now covers Tier 1 customers
6
Readers
Triage agent, and the Support team in their chat
7
Delivered
Live to the triage agentand as a file in the agents' instructions
  1. 5

    A version

    Each release has a number and a list of changes. An agent can stay on one version until its builder is ready.

  2. 6

    Readers

    The agents and people allowed to read it. You can see who reads which version.

  3. 7

    Delivery

    Agents read it live, or as a file in the tools they already use.

Related

A cousin of the data product

Data teams already manage data as a product. A Context Product brings the same habits to the knowledge around the data.

A data product holds tables and metrics, proven by quality checks, for analysis. A Context Product holds rules and definitions, proven by test questions, for agents that act. Both have an owner, a version, a contract and known readers. Data product Tables and metrics Proven by quality checks Built for analysis Both have An owner A version A contract Known readers Context Product Rules and definitions Proven by test questions Built for agents that act

The two work together. A Context Product can name the data product that a figure comes from.

Principles

Seven principles. Seven stars.

We think every Context Product should follow these principles. Use them, share them, and tell us where they are wrong.

  1. 1

    One job per product

    A Context Product serves one job, such as triaging tickets. What it carries follows from that job.

  2. 2

    Every answer has an owner

    A named person or team answers for each rule. When someone leaves, their answers move to a new owner.

  3. 3

    Nothing ships untested

    Each version answers its test questions correctly before it ships. Anyone can check the result.

  4. 4

    The reasons travel with the answer

    Each answer carries its owner, its source and its age. An agent can tell a checked rule from a guess.

  5. 5

    Two answers stay two answers

    When two teams define a word differently, the product says so, and explains how the answers differ.

  6. 6

    Use writes the next test

    Questions that agents could not answer go to their owners. Once answered, they become new test questions.

  7. 7

    Write once, read anywhere

    A Context Product works with any model and any platform. Your context stays yours when you change tools.

How it works

From what your team knows to what your agents use

  1. Step 1

    Write it down

    Start from what you have: a prompt, a policy or a spreadsheet. Agents read it and propose rules for you to check.

  2. Step 2

    Agree on it

    Each rule goes to the person who owns it. Where two teams disagree, both answers stay visible.

  3. Step 3

    Test and ship

    The product answers its test questions, then ships as a new version that your agents read.

  4. Step 4

    Learn from use

    You see what your agents asked. Each question they could not answer goes to its owner as a task.

We are building the tool that does this. One team can start alone, in an afternoon, with no IT project.

Get early access
For technical teams

Where it fits in your stack

Context Products sit between the systems that hold your knowledge and the agents that act on it. Your agent framework still runs your agents. Your catalog stays the reference for your data.

Agents and peopleSupport agentSales assistantChat in Slack or TeamsData platforms
DeliveryMCP server, liveSKILL.mdAGENTS.mdMarkdownSemantic viewsCatalog glossary
Context ProductsVersionedTested before releaseOwnedContract and changelog
Curated context, on one modelDefinitionsRules and thresholdsPoliciesOwnersExceptionsSystems and operations
Your sourcesDocumentsPromptsWarehousesCatalogsRepositoriesChat threads

Context moves up: Context Products read your sources and serve your agents.

  • Live over MCP

    Agents call a product as an MCP server. Each answer carries its owner, source, certification and freshness.

  • Files in your repository

    A product exports as SKILL.md, AGENTS.md or Markdown. A pull request on an exported file becomes a proposal to its owner.

  • Pin a version

    Each release has a semantic version, a changelog and a freshness window. An agent moves when its builder is ready.

  • Explicit states

    When there is no plain answer, the agent gets not defined, provisional or not entitled, with what to do next.

  • Every export reports its loss

    Each destination has a capability profile. An export reports what mapped, what degraded and what was dropped.

  • Access by rule

    Visibility rules decide who reads what. Withheld context is declared, so an agent escalates instead of guessing.

Your catalog is a source. Context Products read warehouses, catalogs and repositories, and can publish approved terms back to a catalog as its glossary.

Early access

Get early access

We are opening to a few teams first. Tell us about your team and the agent you run. We read every answer.

  • You run an AI agent for your team, and it gets things wrong.
  • You lead a team whose rules live in prompts and in people's heads.
  • You build agents and need business context you can pin to a version.

We use your answers to contact you about early access, and for nothing else. Ask us at any time and we delete them.