Microsoft presents: Building and shipping agents in Business Central

Overview

The session from BC TechDays presents agents as a first-party product in Business Central. It walks through creating an agent instance in the client, setting name, language, region and user access, and iterating on profile, permissions and instructions. It then shows testing with task simulation, task templates and the task log, including user intervention steps.

The second half covers shipping. Agents created with instructions in the client cannot run in production, so they are exported and packaged as an extension, with the speaker saying production is supported from version 28. The AL side covers the agent metadata provider enum, the factory and metadata interfaces, and the Copilot capability as a feature gate. The session closes with AI evaluation tools: data-driven, turn-based tests, YAML data sets, an agent test library, credit budgets and run history, available in sandbox only.

Key points

• Agents created with instructions in the client cannot run in production. They must be packaged as an extension and published to the production environment; the speaker says production is supported from version 28.
• Model choice is a dropdown in version 28 (GPT-4, GPT-4.1, with GPT-5 newer). In 27.4 the agent models list is not shown. Models differ in cost, and the speaker expects availability to grow as more models are validated.
• Test agents with task simulation (for email: subject, from, message, attachments) and reusable task templates. The task log shows each run’s steps, tools used, and user intervention requests.
• Export the agent definition from the agents list to collect profile, permissions, instructions, locale and version. Instructions can also be downloaded on their own for an extension.
• Disabling the Copilot capability deactivates all agent instances that use it. Re-enabling it restores them.
• AI evaluation tools run in sandbox only, use your own data sets, and can feed CI/CD. Agent tests are turn-based, use YAML data sets and an AI test type code unit, and depend on the AI development toolkit evaluation extension.
• Budget limits exist at environment and company level. They block new tasks but let running tasks finish, so spend can slightly exceed the limit. Test results show Copilot credits consumed per task; the total tokens column shows only test framework tokens (such as LLM as judge calls), not the agents’ own tokens.

Features

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 “agent” at 1:51
• other “agent models list” at 6:22
• other “task template” at 8:07
• other “task log” at 8:53
• other “agents list” at 13:23
• enum “agent metadata provider” at 19:35
• interface “factory” at 20:17
• interface “agent metadata” at 22:32
• interface “take agent task execution” at 23:18
• page “AI evals suite pages” at 28:07
• codeunit “agent test context” at 40:42
• codeunit “AI test context” at 41:23
• codeunit 4321 “Agent” at 43:22
• page 4316 “Agent List” at 45:25

Not found in BC28-30: enum “agent metadata provider”, interface “factory”, interface “agent metadata”, interface “take agent task execution”, page “AI evals suite pages”, codeunit “agent test context”, codeunit “AI test context”.

Quotes

• 0:04 “Agents are no longer just experiments. They’re becoming a first-party product inside of Business Central.”
• 1:30 “This is all available for you today. You can look at it in the client. Um from 27.4 and newer.”
• 2:37 “Maybe this instance of my agent needs to work in a certain locale. Maybe it’s a US decimal points or it’s a European decimal”
• 4:16 “As you mature the agent, you need to come back and figure out if the agent is seeing too much or not seeing enough”
• 6:22 “you wouldn’t even see this agent models list. That would be for 28 where you have now this drop down”

Disclaimers in the video

• 6:22 subject-to-change: I expect that when we validate more and more models, we will increase the availability here
• 23:50 coming-later: This is a feature that’s coming up. Actually, Quintin here worked on it, and he’s a he’s ready to release …
• 27:53 preview: this is available in sandbox only since this is a testing tool, but you’d be running this before going to production

Presenters (as heard): Blazhe, Stefan, Quintin.