---
title: Agents without context are confident interns.
description: A capable model still gets your business wrong when it lacks context. The five kinds of context an internal agent needs, and where each one lives in a Hub.
canonical_url: https://hubzoid.com/blog/agents-without-context-are-interns
last_updated: 2026-09-27
---

# Agents without context are confident interns.

A capable model still gets your business wrong when it lacks context. The five kinds of context an internal agent needs, and where each one lives in a Hub.

Perspectives · Published 2026-09-25 · Updated 2026-09-27

Picture a new analyst with an excellent degree, a perfect memory and no handover. Nobody has shown them the reporting calendar, introduced them to the warehouse leads or explained why the Leeds numbers always arrive late. On their first morning you ask for a summary of last week's stock movements.

They will produce something. It will be well written. It will be wrong in ways that take an hour to unwind, because it answers the question the way a textbook would rather than the way your business works. An AI agent without context is that analyst.

> The model can reason. What it lacks is everything an insider knows.

This article sorts that missing knowledge into five kinds, and shows where each one lives.

## Five kinds of context at a glance

In Hubzoid, each kind of context has a place in the Hub, the shared set of instructions, notes and tools a team's agents work from. Each row is a section below.

| Context | The question it answers | Where it lives in a Hub |
| --- | --- | --- |
| Identity | What and who are we talking about? | Reference notes for stable facts, tools and connections for live records |
| Process | How do we do it here? | Written skills, one per procedure |
| Decision | What have we already decided, and why? | Dated decision notes with a status |
| Relationship | Who owns this, and who needs to know? | Owner lines in notes, the agent's instructions, and access in the Console |
| Temporal | What is true now, and since when? | Effective dates in notes, a built-in clock, scheduled refreshes |

## A bigger model does not replace a handover

A general model knows how inventory is usually counted, what an accrual is and how a shift handover generally works. It does not know, for example:

- That your team measures slow movers from the last sale date rather than the last receipt.
- That the north warehouse moved its dispatch cut-off in July.
- That the finance director signs off anything above a certain amount.

Those details decide whether an answer is usable, and a larger model does not close this gap.

> A bigger model reasons better about what it is given. The gap is in what it is given.

When people say an agent needs context, they often mean "connect it to our documents". That covers a slice of it. The five kinds below cover the rest.

## Identity context tells the agent what you are talking about

Which warehouse, which supplier, which account code, which product line. Stable facts about the things your business deals with belong in a reference note: a page that lists each site, what it stocks and its known quirks. Live facts, such as today's stock level or an account balance, stay in the system that owns them. The agent reaches them through a tool or a connection to that system, so it never works from a stale copy.

The person asking is part of identity context too. On the surfaces where people sign in, Hubzoid knows who is asking. Tools that only run for people who were granted them check that person's access in code, not a line in the prompt.

## Process context tells it how you do things here

Not how month-end works in general, but how your team does it. Each procedure gets its own written skill: the slow-mover review, the vendor bill check, the handover note. The agent sees a short menu of skill names and descriptions, and reads the full steps only when a task calls for them.

A skill can point to reference notes for the rules it applies. The procedure and the facts can then change separately. A new approval limit is one edit to one note, and every skill that relies on it follows.

## Decision context stops it proposing what you already ruled out

Much of what an insider knows is settled argument. We tried weekly counts at Leeds and went back to daily. We stopped using that carrier. We agreed that shipments held at a customer's request do not count as late.

Without this, an agent will cheerfully propose the option you rejected last quarter. Record each decision as a short, dated note. Here is an invented example of what one might say:

- **Decision.** Leeds counts fast-moving stock daily, not weekly.
- **Decided.** 12 June 2026, by the operations lead and the Leeds supervisor.
- **Reason.** Weekly counts let picking errors go unnoticed for up to six days.
- **Status.** Current, to be reviewed at the end of the year.

Reference notes can sit in subfolders, so all the decisions can live together in one place without any extra setup.

## Relationship context tells it who owns what

Who maintains the stock export, who approves a write-off, who needs to hear about a late shipment before the customer does. Put ownership in the notes it concerns ("Owner: Financial controller") and keep one short page for areas that span several notes. The agent's written instructions can say who it serves and where to send people when a question falls outside its job.

Who may use the agent, and which tools, is a different question. In Hubzoid that belongs in the Console's access settings, where an administrator can see and change it. It does not belong in text the model reads and might be talked out of.

## Temporal context tells it what is true now

The most dangerous answer is one that was right last month. Write effective dates into reference notes, and keep each note describing the current state, with at most a line about what it replaced. Hubzoid also gives the agent a built-in way to check the current date and time in a given time zone, so it can tell whether "the latest figures" are from this morning or last week.

Some notes follow a source, such as a price list exported every week. A scheduled task can refresh those notes from the latest export. While the task is new, it can be set to leave its edits for a person to check before they become part of the Hub's recorded history.

## Build the context before you add agents

It is tempting to start with the agents: a briefing agent, a finance agent, an IT helpdesk agent. Each one then rebuilds the same context slightly differently, and none of them stays current for long.

Invert the order. Take one process, write down the context your best operator carries in their head in a form someone else can maintain, and point one agent at it.

Write down one process → Point one agent at it → Fix what it gets wrong → Add a check → Add the next agent

Context first, then agents, one process at a time.

Each mistake the agent makes shows you a missing note, a vague rule or an undated fact. Fix the Hub, then add a short written check that describes the right answer, so Hubzoid can confirm the fix still holds after later changes. Only then add the next agent.

## Context is the part that carries over

Every later agent, workflow and connected assistant then starts from context that already works. That is the practical meaning of providing context once. The effort you put into explaining your business carries over, instead of being spent again for every new tool.

> The model will keep improving without your help. Your context will not. That is the part worth owning.

For more, read [what a Hub holds](https://hubzoid.com/docs/concepts/hub) or [what belongs in a shared Hub](https://hubzoid.com/blog/what-belongs-in-a-shared-hub).