Have you ever ever questioned what the Ask -> Plan -> Human within the Loop -> Agent workflow truly seems to be like after we use it to construct one thing actual?

In my earlier weblog, Agentic AI in Energy BI and Cloth, Half 4: From Abilities to an Agentic Report Theme Audit Workflow, I defined this workflow and confirmed how abilities, instruments, brokers and Human within the Loop can work collectively in a extra structured AI-assisted course of.

There is a crucial level from that article which is value repeating right here. The Ask -> Plan -> Human within the Loop -> Agent workflow shouldn’t be, by itself, an autonomous agentic AI course of.

Sure, we’re utilizing an AI agent. It could examine the issue, perceive the codebase, create an implementation plan, use instruments and finally change the venture. Nonetheless, the human remains to be actively driving the method between these phases. We ask the preliminary questions, evaluate the plan, problem or appropriate it when required, resolve when it’s adequate, and solely then permit the agent to start the implementation.

I don’t assume there may be something fallacious with that. For a lot of software program and knowledge initiatives, that is at present a smart manner of working with AI. I need the agent to do a lot of the heavy lifting, however I additionally wish to perceive what it plans to do earlier than it begins altering issues. That Human within the Loop checkpoint is vital.

On this weblog, I take that workflow out of the diagrams and use it in an actual venture. We’re constructing a Microsoft Cloth App.

Microsoft Cloth Apps is a brand new code-first expertise for constructing functions straight in Microsoft Cloth. It could create and deploy an software along with its operational knowledge assets inside a Cloth workspace. The appliance on this experiment is an AdventureWorks-style gross sales software with prospects, merchandise, stock, gross sales orders, an approval workflow and fundamental reporting.

This experiment is especially attention-grabbing as a result of I’m not an software developer and I have no idea TypeScript or React. I work primarily with knowledge modelling, Energy BI, DAX, Energy Question and Cloth structure. I do know the enterprise final result I need and perceive the structure and knowledge, however I want GitHub Copilot to assist me work via the appliance implementation.

The appliance was created domestically, deployed to Cloth and examined with a second consumer. It labored, however that’s not an important end result. The extra helpful final result was seeing the way to give an AI coding agent sufficient context, the way to management its scope, and the place the human evaluate factors want to stay.

This publish is one other seed for my wider Agentic AI in Energy BI and Cloth sequence. The sequence started with terminology and psychological fashions, then moved into organising Visible Studio Code and GitHub Copilot safely, utilizing MCP servers with Energy BI, and constructing reusable abilities and agent workflows. This instance sits between these ideas and a completely autonomous workflow. Earlier than we construct autonomous workflows the place an agent decides when to plan, use a talent, name a instrument or take an motion, it helps to see what every stage does in an actual venture.

What This Weblog Will Cowl #

This publish covers:

• what Microsoft Cloth Apps is, and the place it at present makes most sense
• the tenant and native improvement stipulations
• the Ask, Plan, Human within the Loop and Agent workflow I used with GitHub Copilot
• how the Cloth App scaffold, database, software and deployment have been created
• the sensible classes from testing and sharing the appliance
• why Cloth Apps doesn’t merely exchange Energy BI or Energy Apps

The total improvement course of took round 4.5 hours. The YouTube video is far shorter as a result of the lengthy implementation and agent-thinking intervals are fast-forwarded. You may watch the total video right here👇:

Should you’d prefer to take heed to the content material on the go, right here is the AI generated podcase comprised of the content material of this weblog 👇:

What Are Microsoft Cloth Apps? #

Cloth Apps shouldn’t be seen as one other generic website-building platform. It’s a code-first software improvement expertise that runs throughout the Cloth ecosystem. For my part, probably the most attention-grabbing use instances are due to this fact inner operational functions which are near knowledge already managed in Cloth.

Typical examples embody gross sales administration, stock, procurement, approval workflows and different enterprise processes the place customers must create, replace and act on operational knowledge. That is completely different from a conventional Energy BI reporting situation, the place the principle objective is analysing knowledge slightly than altering it.

The excellence issues. Whereas a refined shopping-style demo could make Cloth Apps appear to be a public e-commerce platform, the present worth proposition is stronger for organisation-wide enterprise functions the place Cloth capability, governance, id and knowledge companies are already in place.

Put together the Cloth Setting #

Earlier than creating an software, Cloth Apps have to be enabled within the tenant and the goal workspace have to be backed by Cloth capability.

Allow Cloth Apps #

Within the Cloth portal, open the Admin portal, search tenant settings for “Cloth Apps”, and allow the setting. In a manufacturing tenant, it’s smart to begin with a particular safety group slightly than enabling a brand new functionality for the entire organisation instantly.

The subsequent setting is “Allow Nameless Information Entry for Cloth Apps”. That is meant for public or unauthenticated situations. It might be related for a public product catalogue, timetable or related content material, nevertheless it was not required for this software and remained disabled.

Nameless entry ought to be handled as an structure and safety choice, not as a setting to allow solely as a result of it’s obtainable.

Set up the Native Necessities #

Cloth Apps improvement makes use of the Rayfin SDK and CLI. The official documentation supplies the set up steps, however the native runtime necessities have to be obtainable first.

In my case, the primary try to put in Rayfin failed as a result of npm was not obtainable. After putting in Node.js and npm, the Rayfin CLI put in efficiently and the native surroundings was prepared.

It is a small however sensible level. Agentic improvement doesn’t take away regular improvement dependencies. The agent might help diagnose a lacking dependency or recommend the proper command, nevertheless it nonetheless wants a working native surroundings. It’s a easy factor, however simple to neglect after we begin utilizing these instruments.

Begin With Context, Not Implementation #

The venture began with a principally empty repository and a ZIP file containing a UI mock-up. I used GitHub Copilot Ask mode first, with a intentionally slim instruction: perceive the venture and wait. Don’t create information or generate code.

This created an vital separation between understanding the duty and implementing the duty. An agent requested to do each without delay has to make assumptions in regards to the product, structure and codebase whereas it’s already altering information. That’s not place to begin.

The primary venture context coated the required enterprise areas: prospects, merchandise, stock and gross sales orders. The UI mock-up then gave the agent a concrete visible goal. It was not anticipated to invent the appliance, design each display and implement all technical layers from one imprecise immediate.

Floor the Agent in Official Documentation #

GitHub Copilot didn’t have a devoted Cloth Apps talent loaded for this session. Slightly than counting on mannequin reminiscence for a brand new and shortly altering function, I offered the related Microsoft documentation and requested the agent to review the beneficial venture construction, tooling, native improvement workflow and present limitations.

I used GitHub Copilot’s #codebase command to look the workspace semantically and floor its response within the information that have been truly current. This was useful as a result of Copilot may affirm that no software code existed but as a substitute of assuming an present venture construction.

I used GitHub Copilot’s #fetch command to retrieve the linked Microsoft documentation and add its content material to the present Copilot session.

This doesn’t assure that an agent will make no errors. It does scale back a big supply of uncertainty. For a brand new Microsoft expertise, I favor to make use of official documentation as the first supply of fact, particularly for stipulations, deployment behaviour and supported structure patterns.

Make the Structure Resolution Earlier than Creating Information #

There have been two doable beginning factors:

1. Begin from the UI mock-up and attempt to combine Cloth Apps into it.
2. Create the official Cloth App venture scaffold first, then combine the authorised UI into that construction.

I requested Copilot to check each approaches in opposition to the official steerage and the venture context. It beneficial the second method. This was the appropriate advice as a result of it preserves the venture structure and configuration anticipated by Rayfin, as a substitute of attempting to retrofit Cloth Apps right into a basic frontend venture later.

This choice was made in Ask mode, earlier than the agent was allowed to change the repository.

Add Persistent Mission Guardrails #

As soon as the structure route was clear, I moved to Agent mode. The primary activity was to not generate the appliance. It was to create repository-level Copilot directions.

The instruction file documented the venture objective, the trusted documentation sources, the anticipated workflow, approval boundaries and the behaviour required when info was unclear. That is notably helpful in long-running work as a result of the vital expectations don’t have to be repeated in each immediate.

I used GitHub Copilot’s #create-instructions command to create the instruction file from the context obtainable within the present Copilot session.

And right here is the end result:

The subsequent activity was a Product Necessities Doc (PRD). Copilot inspected the present UI mock-up and created the PRD within the docs folder. I then reviewed it earlier than persevering with.

That evaluate step shouldn’t be non-compulsory. A necessities doc may be nicely formatted and nonetheless include an incorrect assumption. If that assumption reaches implementation, the later price is increased. AI-generated necessities ought to be reviewed in precisely the identical manner as necessities written by an individual.

Use Plan Mode for the Implementation #

After approving the PRD, I switched to Plan mode. The agent was instructed to create a phased implementation plan based mostly on the necessities, repository directions and present venture context.

Every section wanted an goal, the principle modifications, validation steps, dependencies or dangers, and a human approval checkpoint. The plan began with the Cloth App scaffold after which coated the UI integration, database, check knowledge, CRUD operations, reporting and deployment.

The sensible sequence was:

Context -> Documentation -> Structure -> Repository directions -> PRD -> Plan -> Implementation -> Checkpoints -> Testing -> Deployment

Plan mode can’t straight write the plan to the venture folder, so the preliminary plan was saved within the agent’s non-public planning location. Later, after implementation was full, the agent remembered the unique request and supplied to avoid wasting the plan in docs. It was a small element, however a helpful one.

Create the Software and Cloth Assets #

Beginning implementation required the Cloth workspace title and the app merchandise title. I used “AdventureWorks Gross sales Cloth App”. Copilot requested permission to run the required native instructions after which created the preliminary scaffold.

The Cloth portal confirmed new assets showing below the appliance, together with a Cloth SQL database and SQL analytics endpoint.

There was one vital shock. The scaffold course of robotically ran rayfin up, which deployed an actual Cloth merchandise to the workspace. Copilot displayed an Necessary Disclosure explaining this behaviour.

That is precisely why approval boundaries matter. An AI coding agent is not solely altering native information when it could actually run instructions in opposition to cloud companies. The influence of a command have to be understood earlier than it’s authorised. On this instance, I used to be not involved as a result of deploying the modifications was my subsequent deliberate step anyway, so I saved the deployed scaffold and continued with the native improvement server operating. Nonetheless, this clearly demonstrates how harmful and expensive such a mistake may very well be with out correct guardrails. Stating approval necessities in an Directions file shouldn’t be sufficient for real-world agentic workflows; purpose-built Hooks are wanted to implement these boundaries and cease the agent from appearing past its assigned activity. I’ll cowl Hooks within the subsequent article in my Agentic AI sequence.

Implement the Information Mannequin and Software Phases #

The implementation then progressed via the deliberate phases. When the database section began, the Cloth portal confirmed the appliance schema showing within the SQL database.

It’s possible you’ll discover the Todos and Customers tables within the screenshot. They have been created as a result of I unintentionally chosen the Todo template when the agent requested which Cloth Apps venture template to make use of. I didn’t discover the error on the time and solely caught it later whereas modifying the recording for the YouTube video, so I’m calling it out right here to clarify why these unrelated tables seem within the database.

The appliance used an AdventureWorks-style operational mannequin for purchasers, merchandise, stock and orders. Copilot then populated the database with life like check knowledge and carried out the Create, Learn, Replace and Delete operations wanted by the appliance.

The checkpoint course of was vital right here. I can’t confidently evaluate each generated line of TypeScript, however I can evaluate the enterprise course of, the information mannequin, the anticipated consumer behaviour and the Cloth assets that the implementation creates. That is the place subject material information stays important. We don’t must be an knowledgeable in each line of code to evaluate whether or not the answer is smart.

Check the Software Correctly #

As soon as knowledge was current, I examined the first consumer flows. This included searching prospects and orders, looking merchandise, checking stock, and reviewing the order-approval course of.

The approval workflow allowed a pending order to be opened, reviewed, annotated and authorised. Its standing then moved to processing.

The reviews web page was initially empty. I requested Copilot to implement it as a follow-up section. The ensuing reporting web page is a helpful place to begin, though it’s not the completed reporting expertise I might anticipate in a manufacturing software.

This is a crucial expectation to set with AI-generated functions. A quick first implementation shouldn’t be the identical as a completed resolution. It nonetheless wants iteration, testing, efficiency checks, accessibility evaluate, safety evaluate and consumer suggestions.

Deploy and Share the App #

After native testing, I requested Copilot to run the deployment. The appliance grew to become obtainable via a hosted Cloth App URL.

This modifications the end-user expertise. Customers don’t must navigate via the Cloth portal or workspace to open the appliance. They’ll use the hosted URL straight, though they nonetheless must authenticate with their Microsoft Entra ID account.

For sharing, I used the Cloth Handle entry dialog and granted one other consumer Learn and Execute permissions. I didn’t grant edit or re-sharing permissions.

The consumer didn’t obtain an invite electronic mail on this check. The permissions have been appropriate, however I nonetheless wanted to supply the hosted software URL manually. That is value validating in every organisation slightly than assuming the sharing expertise will work like an ordinary Microsoft 365 invitation movement.

A Helpful Bug Discovered Throughout Testing #

The second consumer created a brand new order efficiently. The client, merchandise, portions and totals have been all recorded. Nonetheless, the salesperson related to the brand new order was incorrect.

I intentionally left the problem seen within the video as a result of it demonstrates a easy reality: AI-generated code remains to be code. It could include bugs. It wants purposeful testing, necessities validation and, for an actual venture, the traditional engineering practices round testing, evaluate and launch administration.

The appliance was spectacular for a primary construct, however no accountable staff ought to interpret a working demo as proof of manufacturing readiness. I believe that is the place we must be life like about AI-generated functions.

Will Cloth Apps Exchange Energy BI? #

No, not within the easy sense individuals typically recommend.

Cloth Apps can assist analytical experiences and might show charts, however displaying a chart doesn’t make an software a Energy BI alternative. Cloth Apps is code-first. Energy BI report authoring is designed to be accessible to analysts and enterprise customers via a low-code interface.

Energy BI modelling may be technically demanding. Severe semantic fashions nonetheless require DAX, Energy Question, knowledge modelling and generally supply management and superior improvement tooling. Even so, authoring a report stays a special expertise from constructing and sustaining a customized software.

There will likely be instances the place an organisation wants a completely customized analytical consumer expertise and Cloth Apps is a smart choice. In lots of reporting situations, Energy BI stays the clearer selection. The 2 merchandise have overlap, however they remedy completely different issues.

Will Cloth Apps Exchange Energy Apps? #

The identical conclusion applies to Energy Apps. Energy Apps is a low-code improvement platform. Cloth Apps is a code-first expertise.

There will likely be overlap in inner enterprise functions, however the selection ought to rely on the appliance necessities, operational knowledge structure, present abilities, governance mannequin, required consumer expertise and long-term assist mannequin. Product overlap doesn’t imply one product replaces the opposite.

An Structure Query That Wants Extra Consideration #

The appliance didn’t behave like a easy consumer interface positioned on high of an present transactional database. The Cloth App venture was concerned in defining and managing its operational knowledge construction.

This has vital implications. If an organisation already has a fastidiously designed transactional system and expects Cloth Apps to be solely one other entrance finish for that database, it must validate the present Cloth Apps structure and supported patterns fastidiously. That query is extra vital than the visible high quality of a demo software.

The subsequent space I wish to discover is the analytical structure. What occurs when a Cloth App must work with knowledge and enterprise logic that already exists in Cloth and Energy BI? What permissions does the appliance want? What permissions do finish customers want? How does sharing work together with semantic mannequin permissions and safety?

These are the questions that decide whether or not Cloth Apps can be utilized nicely in an actual organisation.

Within the subsequent video, I proceed the Cloth Apps journey and go deeper into what this new expertise actually means for Energy BI, Energy Apps, and enterprise customers. I look past the demo and break down the precise structure, how Cloth Apps creates and manages the information layer, how GitHub Copilot might help construct an actual app, and the place this matches in the way forward for analytics and inner enterprise functions. So keep tuned for that one.

Last Ideas #

This experiment produced a working AdventureWorks-style software in round 4 and a half hours, regardless of software improvement not being my background. GitHub Copilot helped create the venture construction, necessities, plan, database, consumer interface, knowledge operations, reporting and deployment.

The vital lesson shouldn’t be that an agent can generate an software from one immediate. It can’t, a minimum of not reliably. The higher lesson is that area consultants can use agentic instruments to discover and construct options that might in any other case be outdoors their technical attain, offered they maintain the method grounded in documentation, structure selections, approval factors and testing. This half, for me, is the extra attention-grabbing change.

The sensible method is straightforward: begin with context, plan earlier than implementation, maintain the human within the loop, and check the end result as significantly as every other software.

Should you discovered this handy, you’ll be able to observe me on LinkedIn, subscribe to BI Perception on YouTube, or observe the BI Perception podcast on Spotify. I additionally share updates on Bluesky and X, protecting Energy BI, Microsoft Cloth, AI, and the real-world knowledge and analytics issues I come throughout alongside the way in which.

Associated