Give a draft an accountable owner
Use existing HVE research, planning, writing, and review specialists. Ground their work in your approved material; the package does not ship custom sales or other domain-expert agents.
See the role-to-work mapping →Business work. Native MAF execution.
A squad.
Your app.
Real evidence.
Help your application's users prepare a proposal response, compare customer solution options, or draft a knowledge article. Embed released HVE Squad roles to research, plan, write, and review with evidence and human approval.
Your IChatClient. Your case data. No separate squad server.
One reviewed business draft.NET / MAF
- 01 · Research Approved case inputs
- 02 · Plan + details A response outline
- Human approval The app's decision UI
- 03 · Produce One draft owner
- 04 · Review Evidence + open gaps
Conditional intake and council add gates when required.
Native, not a proxyMicrosoft.Agents.AI 1.17.0 on .NET 10
Released, not branch-headStable release → tag → commit → APM graph
Preview, not publishedUse a project reference or local package feed
Why HVE Squad inside MAF?
Bring the method into the application.
This is a .NET library intended for NuGet consumption inside your AI application, not a Copilot plugin or a tool for developing the application itself. MAF supplies agent execution; this integration adds released roles, an approved plan, a single producing owner, review, and evidence-backed state. The preview is available from source or a local feed; see how maintainers publish the package.
Keep the application in charge
Your app owns identity, document rights, case isolation, the model provider, and the approval UI. A drafted recommendation cannot authorize a price, change a customer record, or send itself.
Understand application responsibilities →Return evidence, not just an answer
Expose the draft alongside research, plan, review, and an explicit run status. The deterministic Scribe records actual dispatches and hashes; those receipts are not proof that every business claim is correct.
Read the result contract →Application features that benefit from review.
RFP response workbench
A sales user submits an RFP and approved capabilities in your portal. Return one cited Markdown response and gaps for bid-manager review, never an automatically submitted bid.
Worked application example →Customer solution advice
An internal adviser compares options against customer needs and approved catalog evidence. Return a decision brief, not generated code or a configured customer system.
Solution-advisory portal →Product feedback synthesis
A product manager turns existing approved feedback into an evidence-linked discovery brief and roadmap proposal. No invented interviews or automatic tracker changes.
Feedback application →Knowledge operations
A knowledge steward requests an article grounded in approved policy and process text. Return a reviewed draft to its owner, not an automatically published policy.
Knowledge application →A clear boundary is a feature.
Interactive. Single squad. One producing-role owner per turn.
These are advisory document workflows using HVE specialists, not turnkey domain products. Input intake uses the released PRD Quality Reviewer and may block non-PRD material; assess that fit before offering an application feature.
This preview does not implement federation, Watch Mode, autonomous/autopilot loops, discovery interviews, or multi-owner fan-out. Native child dispatch is real delegation, but it is not federation.
Why this scope was chosen, and what federation would require →The host supplies document ingestion, retrieval, approval transport, and secure output delivery. Native Markdown and evidence are not a Word/PDF service, pricing engine, or compliance proof. Completed does not mean externally sent or independently validated. For commodity ticket routing or simple summarization, plain MAF is often enough.
From documentation to an application job
Start with approved text. Return one reviewed draft.
Inspect the release without model credentials. Bind your IChatClient, provision an isolated case directory, and connect real human callbacks. Keep the first job free of external writes. No code checkout is required for a business case.