Agentic AI Solution Accelerator¶
A GitHub template that Microsoft partners clone to deliver a customer-specific agentic AI solution — live in days, not months. The full engagement motion (discovery → UAT → handover → measure) is weeks, and walked step-by-step below.
Flagship scenario: Sales Research & Personalised Outreach — a supervisor agent routes a research request across specialist workers (Account Researcher, ICP/Fit Analyst, Competitive Context, Outreach Personaliser) and returns a grounded, citeable sales brief with a CRM-ready outreach draft. Human-in-the-loop gates every CRM write and every email send.
Stack: Microsoft Agent Framework · Microsoft Foundry · Azure AI Search · Managed Identity · Key Vault · Container Apps · Application Insights · azd for infra.
How this site is organised¶
The accelerator is delivered as a linear walkthrough in two tracks. Read in order; each step ends with a single Continue → link to the next.
- Get ready (3 steps, do once) — orient yourself, install your tools, rehearse in your own sandbox.
- Deliver to a customer (7 steps, repeat per engagement) — clone, discover, scaffold, provision, iterate, hand over, operate.
Roles are skim guidance, not separate paths:
| Role | Read deeply | Skim |
|---|---|---|
| Delivery lead | Get oriented, Discover with the customer, UAT & handover, Operate (Day 2) | The engineering steps |
| Partner engineer | Set up your machine, Rehearse in a sandbox, Clone through Iterate & evaluate | Get oriented |
| Solo partner | All 10 steps | — |
Joining mid-engagement?¶
Jump in at the step that matches your current state.
| If this is true… | Start at |
|---|---|
| You haven't cloned the customer repo yet | 4. Clone for the customer |
| Repo is cloned, no brief yet | 5. Discover with the customer |
| Brief is filled in the repo, no scaffold yet | 6. Scaffold from the brief |
| Code scaffolded, not provisioned in customer Azure | 7. Provision the customer's Azure |
| Provisioned, you're customising and running evals | 8. Iterate & evaluate |
| Evals green, heading into UAT | 9. UAT & handover |
| Already in production | 10. Operate (Day 2) |
Reference material¶
Deep dives that sit outside the walkthrough — open them when a step sends you there.
Common tasks (jump straight to the chatmode or page you need)¶
| If you want to… | Go to |
|---|---|
| See which chatmode runs at which step | Chatmodes overview |
| Add a side-effect tool with HITL baked in | /add-tool (used in step 8) |
| Add a specialist worker agent | /add-worker-agent (used in step 8) |
| Scaffold the customer's repo from a filled brief | /scaffold-from-brief (used in step 6) |
| Onboard a new Azure environment (dev/UAT/prod) | /deploy-to-env (used in step 7) |
| Pick a landing-zone tier (standalone / AVM / ALZ-integrated) | /configure-landing-zone (used in step 7) |
| Preflight a change against lint / eval / deploy gates | /explain-change (used in step 8) |
| Re-author the scenario as single-agent or chat-with-actioning | /switch-to-variant (used in step 8) |
Deep dives¶
- Architecture & governance — Reference architecture · Azure AI landing zone · WAF alignment · Responsible AI
- Engineering reference — Agent specs · Tool catalog · Version matrix
- Frontend starter —
patterns/sales-research-frontend/ - Reference scenarios — Customer service actioning · RFP response
- Delivery context — Partner playbook (narrative companion to the walkthrough) · Solution brief guide · Handover template · Workflow map
GitHub repo · Contributing · Support
Returning for the next customer?¶
Track 1 (Get ready) stays done. Skip straight to 4. Clone for the customer and run through Track 2 again with the new customer's short-name. The accelerator is built to support multiple customer engagements from one prepared workstation.
Quick preflight before the next engagement:
-
gh auth statussucceeds and you can see the customer's GitHub org. -
az account listincludes the customer's tenant; partner has rights in their target subscription. -
accelerator-lint.pyis green on the latest commit of the template (main). - You've reviewed any version-matrix updates from the weekly CI run.
Glossary¶
Acronyms used across these docs, in one place.
| Term | Meaning |
|---|---|
| ALZ | Azure Landing Zone — Microsoft's reference architecture for an enterprise Azure tenancy (hub vNet, private DNS, MG hierarchy, policy). |
| AVM | Azure Verified Modules — Microsoft-maintained Bicep/Terraform modules following Azure best practices. |
| BYO | Bring Your Own — partner supplies the Azure subscription / tenant / identity, accelerator deploys into it. |
| CCoE | Cloud Center of Excellence — the customer team that owns Azure platform standards, ALZ, policy. |
| HITL | Human-in-the-Loop — required human approval gate before any side-effect tool call (CRM write, email send, etc.). |
| MG | Management Group — Azure's tenancy-scoped policy + RBAC container above subscriptions. |
| OIDC | OpenID Connect — token-based federated auth; used here so GitHub Actions can deploy to Azure without a service-principal secret. |
| RAG | Retrieval-Augmented Generation — pattern of grounding LLM responses in indexed sources (here, Azure AI Search). |
| RAI | Responsible AI — Microsoft's safety practice (content filters, redteam evals, PII handling, transparency). |
| RBAC | Role-Based Access Control — Azure's identity-based permission model. |
| SSE | Server-Sent Events — one-way streaming HTTP protocol the scenario endpoint uses to push agent progress to clients. |
| WAF | Well-Architected Framework — Microsoft's five-pillar architecture review (reliability, security, cost, op-ex, performance). |