---
id: video/1IIlArcHpOY
type: video
title: "What's New: Working with Instructions for Agents in Business Central"
summary: "Writing instructions for Business Central agents: role-based prompting, a guidelines section, and short process steps. Shown with an Accounts Receivable agent that handles a sales return, reads documents through Azure Document Intelligence, creates a credit memo and drafts an email. Also covers agent memory, user interaction tools and page-specific instructions."
tier: official
language: en
tags:
  - agent instructions
  - role-based prompting
  - guidelines
  - document intelligence
  - agent review experience
  - credit memo
  - sales returns
  - prompt engineering
  - task execution
  - permissions
  - user interaction tools
  - agent memory
system: copilot
review:
  state: reviewed
  by: opus
  at: "2026-10-07T22:51:21.282Z"
  flags: []
generated:
  at: "2026-10-09T00:29:02.995Z"
  pipeline: 0.2.0
  prompts:
    extract-video: 1
    summarize-video: 2
  input_hash: 6e43abefb582523914e6618e43535f7b52dcb15f88ca7a602980af5fee80c22e
evidence:
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=73s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 73
    quote: Today because the models became much better. The rule that we are using is that less is more.
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=84s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 84
    quote: You do not need to outline every specific step that the agent needs to do. It is sufficient to outline the most important steps
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=115s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 115
    quote: instructions are a part of the bigger prompt so everything that you write the platform is going to take it and
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=251s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 251
    quote: based prompting it helps setting the context for the agent because if it knows who it is and which task it is supposed
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=464s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 464
    quote: Agent is going to be able to read this document because platform is able to send PDFs and images to Azure document intelligence and
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=519s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 519
    quote: The platform is only extracting text. It is not able to do the machine vision. So it will know that there is a scratch
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=804s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 804
    quote: currently within the current platform it is not possible that this happens automatically. You would need to use the SDK if you want to
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=879s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 879
    quote: instructions are always going to win over tasks because platform is giving a much higher priority to the instructions compared to the tasks
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=893s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 893
    quote: you should outline the process description in the instructions because there is no need and it's not a good practice that you repeat a
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1043s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1043
    quote: there is no silver bullet or a rule that fits all. Try it out. see what works and if adding the additional instructions is
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1215s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1215
    quote: when it uses the response tool the task is going to stop until the new message arrives
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1264s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1264
    quote: If you would like to store the state of the page then you need to instruct it specifically which fields and which data it
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1345s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1345
    quote: If you add the parenthesis like this and you specify page ID, then these page specific instructions will be available to the agent only
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1410s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1410
    quote: instructions are necessary and keep in mind they are part of a larger prompt. So platform is going to protect you from injection attacks
  - kind: video
    url: https://www.youtube.com/watch?v=1IIlArcHpOY&t=1427s
    title: "What's New: Working with Instructions for Agents in Business Central"
    date: "2026-02-27T15:12:59.000Z"
    commit: null
    t: 1427
    quote: Try to write as little as possible of instructions to get the task executed with high accuracy
links:
  learn: []
  objects: []
  features:
    - feature/573366
  topics: []
  localizations: []
  videos: []
  posts: []
  guidelines: []
video_id: 1IIlArcHpOY
channel: yt-microsoft
source_name: Microsoft Dynamics 365 Business Central (YouTube)
url: https://www.youtube.com/watch?v=1IIlArcHpOY
published_at: "2026-02-27T15:12:59.000Z"
duration_s: 1517
captions: full
audience:
  - developer
  - functional consultant
  - administrator
chapters:
  - t: 0
    title: Introduction and Agent Instructions Strategy
  - t: 156
    title: Creating an Accounts Receivable Agent
  - t: 237
    title: Structuring Instructions with Guidelines and Business Processes
  - t: 352
    title: Sales Returns Task Setup and Document Processing
  - t: 464
    title: Agent Execution and Document Intelligence
  - t: 605
    title: Agent Review Experience and Credit Memo Creation
  - t: 742
    title: Email Composition and Task Completion
  - t: 823
    title: Instructions versus task messages in agent design
  - t: 928
    title: Timeline defaults and agent subtasks
  - t: 1009
    title: Agent scope and task chaining
  - t: 1087
    title: User interaction tools overview
  - t: 1194
    title: Agent memory and data retention
  - t: 1307
    title: Advanced tools and page-specific instructions
  - t: 1384
    title: Best practices and optimization guidance
features:
  - name: Agent Instructions Framework
    status: unclear
    t: 5
    verified: false
    status_source: video
  - name: Accounts Receivable Agent
    status: unclear
    t: 156
    verified: false
    status_source: video
  - name: PDF and Image Document Intelligence
    status: unclear
    t: 464
    verified: false
    status_source: video
  - name: Agent Review Experience
    status: ga
    t: 650
    verified: false
    status_source: roadmap
    roadmap_ids:
      - "573366"
  - name: Agent Task Execution with Real-time Monitoring
    status: unclear
    t: 546
    verified: false
    status_source: video
  - name: Credit Memo Generation from Agent Instructions
    status: unclear
    t: 650
    verified: false
    status_source: video
  - name: Email Composition by Agent
    status: unclear
    t: 766
    verified: false
    status_source: video
  - name: Instructions Priority Over Tasks
    status: unclear
    t: 837
    verified: false
    status_source: video
  - name: Timeline substeps based on instructions
    status: unclear
    t: 950
    verified: false
    status_source: video
  - name: Task chaining in agent workflows
    status: unclear
    t: 1059
    verified: false
    status_source: video
  - name: User intervention tool
    status: unclear
    t: 1103
    verified: false
    status_source: video
  - name: User review tool
    status: unclear
    t: 1131
    verified: false
    status_source: video
  - name: Response tool for outgoing messages
    status: unclear
    t: 1144
    verified: false
    status_source: video
  - name: Automatic agent memory and history
    status: unclear
    t: 1229
    verified: false
    status_source: video
  - name: Agent memory field storage
    status: unclear
    t: 1253
    verified: false
    status_source: video
  - name: Page-specific instructions syntax
    status: unclear
    t: 1332
    verified: false
    status_source: video
  - name: Instructions history and versioning
    status: unclear
    t: 1384
    verified: false
    status_source: video
  - name: Bulk download of instructions
    status: unclear
    t: 1399
    verified: false
    status_source: video
objects_mentioned:
  - other accounts receivable administrator
  - other sales order agent
  - other Paris guest chair
  - other Berlin guest chair
  - page troubleshooting page
quotes:
  - t: 73
    text: Today because the models became much better. The rule that we are using is that less is more.
    check: exact
  - t: 84
    text: You do not need to outline every specific step that the agent needs to do. It is sufficient to outline the most important steps
    check: exact
  - t: 115
    text: instructions are a part of the bigger prompt so everything that you write the platform is going to take it and
    check: fuzzy
  - t: 251
    text: based prompting it helps setting the context for the agent because if it knows who it is and which task it is supposed
    check: fuzzy
  - t: 464
    text: Agent is going to be able to read this document because platform is able to send PDFs and images to Azure document intelligence and
    check: exact
  - t: 519
    text: The platform is only extracting text. It is not able to do the machine vision. So it will know that there is a scratch
    check: exact
  - t: 804
    text: currently within the current platform it is not possible that this happens automatically. You would need to use the SDK if you want to
    check: exact
  - t: 879
    text: instructions are always going to win over tasks because platform is giving a much higher priority to the instructions compared to the tasks
    check: exact
  - t: 893
    text: you should outline the process description in the instructions because there is no need and it's not a good practice that you repeat a
    check: exact
  - t: 1043
    text: there is no silver bullet or a rule that fits all. Try it out. see what works and if adding the additional instructions is
    check: exact
  - t: 1215
    text: when it uses the response tool the task is going to stop until the new message arrives
    check: exact
  - t: 1264
    text: If you would like to store the state of the page then you need to instruct it specifically which fields and which data it
    check: exact
  - t: 1345
    text: If you add the parenthesis like this and you specify page ID, then these page specific instructions will be available to the agent only
    check: exact
  - t: 1410
    text: instructions are necessary and keep in mind they are part of a larger prompt. So platform is going to protect you from injection attacks
    check: exact
  - t: 1427
    text: Try to write as little as possible of instructions to get the task executed with high accuracy
    check: exact
---

# What's New: Working with Instructions for Agents in Business Central

> Writing instructions for Business Central agents: role-based prompting, a guidelines section, and short process steps. Shown with an Accounts Receivable agent that handles a sales return, reads documents through Azure Document Intelligence, creates a credit memo and drafts an email. Also covers agent memory, user interaction tools and page-specific instructions.

[Watch on YouTube](https://www.youtube.com/watch?v=1IIlArcHpOY) · Microsoft Dynamics 365 Business Central (YouTube) · 2026-02-27 · 25:17 · tier official · reviewed (checked by Opus)

## Overview

The video explains how to write instructions for agents in Business Central. The guidance is "less is more": outline only the most important steps, use role-based prompting to set context, and keep the process description in the instructions rather than repeating it in each task. Instructions are part of a larger platform prompt and take priority over task messages.

A demo Accounts Receivable agent processes a sales return complaint about damaged items. The platform sends PDFs and images to Azure Document Intelligence, which extracts text only. The agent fills a credit memo, the user reviews it and continues the task, and the agent composes an email. The video also covers timeline substeps, task chaining, user intervention and review tools, the response tool, agent memory, page-specific instructions, instruction history and bulk download.

## Key points

- Keep instructions short: outline the most important steps, not every step. If adding instructions lowers accuracy, consider splitting the agent into several agents. Tasks that are different enough can share one agent.
- Instructions take much higher priority than task messages, so put the process description in the instructions and do not repeat it in tasks. Use task messages for what varies between tasks.
- Document Intelligence extracts text from PDFs and images only. It does no machine vision, so it cannot identify visual patterns on a photo.
- The agent review experience shows agent-filled fields with explanations and lets the user submit feedback. In the demo, guidelines tell the agent not to post draft credit memos before user review.
- In the current platform, the drafted email cannot be sent automatically. Use the SDK to actually send it.
- Agent history is kept automatically, but page state is not. Tell the agent which fields to memorize; they are stored as key-value pairs, viewable on the troubleshooting page.
- Page-specific instructions use parenthesis notation with the page ID and apply only when the agent opens that page. Instruction edit history allows reverting, and several instructions can be downloaded as a zip.

## Chapters

- [0:00](https://www.youtube.com/watch?v=1IIlArcHpOY&t=0s) Introduction and Agent Instructions Strategy
- [2:36](https://www.youtube.com/watch?v=1IIlArcHpOY&t=156s) Creating an Accounts Receivable Agent
- [3:57](https://www.youtube.com/watch?v=1IIlArcHpOY&t=237s) Structuring Instructions with Guidelines and Business Processes
- [5:52](https://www.youtube.com/watch?v=1IIlArcHpOY&t=352s) Sales Returns Task Setup and Document Processing
- [7:44](https://www.youtube.com/watch?v=1IIlArcHpOY&t=464s) Agent Execution and Document Intelligence
- [10:05](https://www.youtube.com/watch?v=1IIlArcHpOY&t=605s) Agent Review Experience and Credit Memo Creation
- [12:22](https://www.youtube.com/watch?v=1IIlArcHpOY&t=742s) Email Composition and Task Completion
- [13:43](https://www.youtube.com/watch?v=1IIlArcHpOY&t=823s) Instructions versus task messages in agent design
- [15:28](https://www.youtube.com/watch?v=1IIlArcHpOY&t=928s) Timeline defaults and agent subtasks
- [16:49](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1009s) Agent scope and task chaining
- [18:07](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1087s) User interaction tools overview
- [19:54](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1194s) Agent memory and data retention
- [21:47](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1307s) Advanced tools and page-specific instructions
- [23:04](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1384s) Best practices and optimization guidance

## Features

| Feature | Status | At |
|---|---|---|
| Agent Instructions Framework | status not stated, demoed | [0:05](https://www.youtube.com/watch?v=1IIlArcHpOY&t=5s) |
| Accounts Receivable Agent | status not stated, demoed | [2:36](https://www.youtube.com/watch?v=1IIlArcHpOY&t=156s) |
| PDF and Image Document Intelligence | status not stated, demoed | [7:44](https://www.youtube.com/watch?v=1IIlArcHpOY&t=464s) |
| Agent Review Experience | generally available (roadmap [Review content generated by agents directly on pages](../features/573366.md)), demoed | [10:50](https://www.youtube.com/watch?v=1IIlArcHpOY&t=650s) |
| Agent Task Execution with Real-time Monitoring | status not stated, demoed | [9:06](https://www.youtube.com/watch?v=1IIlArcHpOY&t=546s) |
| Credit Memo Generation from Agent Instructions | status not stated, demoed | [10:50](https://www.youtube.com/watch?v=1IIlArcHpOY&t=650s) |
| Email Composition by Agent | status not stated, demoed | [12:46](https://www.youtube.com/watch?v=1IIlArcHpOY&t=766s) |
| Instructions Priority Over Tasks | status not stated, demoed | [13:57](https://www.youtube.com/watch?v=1IIlArcHpOY&t=837s) |
| Timeline substeps based on instructions | status not stated, demoed | [15:50](https://www.youtube.com/watch?v=1IIlArcHpOY&t=950s) |
| Task chaining in agent workflows | status not stated, demoed | [17:39](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1059s) |
| User intervention tool | status not stated, demoed | [18:23](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1103s) |
| User review tool | status not stated, demoed | [18:51](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1131s) |
| Response tool for outgoing messages | status not stated, demoed | [19:04](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1144s) |
| Automatic agent memory and history | status not stated, demoed | [20:29](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1229s) |
| Agent memory field storage | status not stated, demoed | [20:53](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1253s) |
| Page-specific instructions syntax | status not stated, demoed | [22:12](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1332s) |
| Instructions history and versioning | status not stated, demoed | [23:04](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1384s) |
| Bulk download of instructions | status not stated, demoed | [23:19](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1399s) |

A status with a roadmap link comes from the Microsoft 365 roadmap feature this part of the video covers (matched by Haiku; links Opus dropped are not used); other statuses need a status word in the video itself.

## AL objects mentioned

As heard in the captions. A name that matches one object page by exact type and name links to it; the others stay as named.

- other "accounts receivable administrator" at [3:21](https://www.youtube.com/watch?v=1IIlArcHpOY&t=201s)
- other "sales order agent" at [4:58](https://www.youtube.com/watch?v=1IIlArcHpOY&t=298s)
- other "Paris guest chair" at [11:52](https://www.youtube.com/watch?v=1IIlArcHpOY&t=712s)
- other "Berlin guest chair" at [11:52](https://www.youtube.com/watch?v=1IIlArcHpOY&t=712s)
- page "troubleshooting page" at [21:22](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1282s)

Not found in BC28-30: page "troubleshooting page".

## Quotes

- [1:13](https://www.youtube.com/watch?v=1IIlArcHpOY&t=73s) "Today because the models became much better. The rule that we are using is that less is more."
- [1:24](https://www.youtube.com/watch?v=1IIlArcHpOY&t=84s) "You do not need to outline every specific step that the agent needs to do. It is sufficient to outline the most important steps"
- [1:55](https://www.youtube.com/watch?v=1IIlArcHpOY&t=115s) "instructions are a part of the bigger prompt so everything that you write the platform is going to take it and"
- [4:11](https://www.youtube.com/watch?v=1IIlArcHpOY&t=251s) "based prompting it helps setting the context for the agent because if it knows who it is and which task it is supposed"
- [7:44](https://www.youtube.com/watch?v=1IIlArcHpOY&t=464s) "Agent is going to be able to read this document because platform is able to send PDFs and images to Azure document intelligence and"
- [8:39](https://www.youtube.com/watch?v=1IIlArcHpOY&t=519s) "The platform is only extracting text. It is not able to do the machine vision. So it will know that there is a scratch"
- [13:24](https://www.youtube.com/watch?v=1IIlArcHpOY&t=804s) "currently within the current platform it is not possible that this happens automatically. You would need to use the SDK if you want to"
- [14:39](https://www.youtube.com/watch?v=1IIlArcHpOY&t=879s) "instructions are always going to win over tasks because platform is giving a much higher priority to the instructions compared to the tasks"
- [14:53](https://www.youtube.com/watch?v=1IIlArcHpOY&t=893s) "you should outline the process description in the instructions because there is no need and it's not a good practice that you repeat a"
- [17:23](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1043s) "there is no silver bullet or a rule that fits all. Try it out. see what works and if adding the additional instructions is"
- [20:15](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1215s) "when it uses the response tool the task is going to stop until the new message arrives"
- [21:04](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1264s) "If you would like to store the state of the page then you need to instruct it specifically which fields and which data it"
- [22:25](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1345s) "If you add the parenthesis like this and you specify page ID, then these page specific instructions will be available to the agent only"
- [23:30](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1410s) "instructions are necessary and keep in mind they are part of a larger prompt. So platform is going to protect you from injection attacks"
- [23:47](https://www.youtube.com/watch?v=1IIlArcHpOY&t=1427s) "Try to write as little as possible of instructions to get the task executed with high accuracy"

## Disclaimers in the video

- [13:24](https://www.youtube.com/watch?v=1IIlArcHpOY&t=804s) not-in-this-release: currently within the current platform it is not possible that this happens automatically. You would need to use the SDK if you want to actually send the email

Presenters (as heard): Nicolola Kukria.