A GitHub template Microsoft partners clone to deliver a customer-specific agentic AI deployment in days. Best read as a rendered site at https://azure-samples.github.io/agentic-ai-solution-accelerator/; this README is what you see inside a clone. Same markdown either way.
Full engagement motion (discovery → UAT → handover → measure) is weeks, and documented below.
Flagship scenario: Sales Research & Personalized Outreach — a supervisor agent routes a research request across specialist workers (Account Researcher, ICP/Fit Analyst, Competitive Context, Outreach Personalizer) 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.
Adoption model: gh repo create --template → Copilot-guided discovery → azd up → iterate in VS Code → merge through CI gates.
👉 Scan the full workflow first: docs/partner-workflow.md — one-page visual of all 7 stages (discover → scaffold → provision → iterate → UAT → handover → measure) across the three responsibilities. Use it to orient yourself, then come back here and open your lane below.
Doc precedence: when two sources conflict, the more specific one wins — chatmode > playbook / QUICKSTART > this README. Pages and this README render the same markdown, so "the doc you're reading" is never the conflict; it's always the linked chatmode or playbook that supersedes. Full chain: Reference material.
- Start with:
docs/partner-playbook.md— end-to-end 7-stage motion, SOW guidance, "what good looks like" per stage - Then run:
/delivery-guidein Copilot Chat for a guided pass through the motion - Also use:
docs/discovery/how-to-use.md(sequences the 5 discovery artifacts) ·docs/handover/handover-packet-template.md(engagement-specific handover template) - Customer already gave you a PRD/BRD/spec? Run
/ingest-prdto pre-draft the brief, then/discover-scenariogap-fills the TBDs. Full flow insidehow-to-use.md. - ✅ Done when: customer sponsor signs off at UAT (Stage 5), handover packet is delivered with a named owner and date (Stage 6), and the first monthly value review is on the calendar (Stage 7).
- Start with:
QUICKSTART.md— clone → discover → scaffold → preflight (/configure-landing-zone+/deploy-to-env) →azd up→ iterate - Then run:
/scaffold-from-briefonce a solution brief exists - Also use:
docs/getting-started/setup-and-prereqs.md(authoritative setup, prereqs,azd uptroubleshooting) ·docs/enablement/hands-on-lab.md(7-lab sandbox rehearsal — strongly recommended before your first customer-facing deployment) - ✅ Done when: acceptance evals (quality + redteam) pass in the customer's environment and the handover artifacts — repo access, runbook, approver rota, killswitch drill notes — are delivered to customer ops.
- Primary: Your engagement-specific handover packet (partner delivers at handover — Stage 6)
- Fallback:
docs/customer-runbook.md— generic day-2 ops (monitoring, killswitch, evals, model swap, secret rotation, incidents). Partner packet wins on conflict. - ✅ Done when (handover accepted): alerts route to your on-call, HITL approver rota is current, killswitch + secret-rotation drills have been run once, and you know which partner contact handles expansion requests. Day-2 ops is steady-state, not a finish line.
Wearing multiple hats at a small partner? The lanes above are responsibilities, not required job titles. Solo partner: run the Lead lane top-to-bottom through Stage 1; drop into the Engineer lane at Stage 2 (scaffold → provision → iterate); return to the Lead lane at Stage 5 (UAT) through Stage 7. Customer ops is always the customer's lane.
Full doc precedence when guidance disagrees (click to expand)
Chatmodes in .github/chatmodes/ (they drive the executable surface) → docs/partner-playbook.md (delivery motion) and docs/getting-started/setup-and-prereqs.md (setup mechanics) → this README. The engagement-specific handover packet supersedes the generic docs/customer-runbook.md for the customer ops lane.
Architecture · WAF alignment · Responsible AI · Azure AI Landing Zone
single-agent · chat-with-actioning · sales-research-frontend (reference UI)
customer-service-actioning · rfp-response
Foundry tool catalog · Agent specs · SDK version matrix
Full code + infra directory tree (click to expand)
agentic-ai-solution-accelerator/
├── accelerator.yaml engagement manifest — scenario contract + acceptance + controls + KPIs
├── src/
│ ├── main.py scenario-agnostic FastAPI; mounts the scenario endpoint from manifest
│ ├── workflow/ framework: BaseWorkflow Protocol + scenario registry (load_scenario)
│ ├── retrieval/ generic SearchRetriever(index_name) against Azure AI Search
│ ├── tools/ HITL-gated side-effect tools (CRM write, email send)
│ ├── accelerator_baseline/ partner-owned primitives: telemetry, HITL, killswitch, evals, cost
│ └── scenarios/ scenario instances loaded via manifest
│ └── sales_research/ flagship: schema, workflow factory, retrieval schema
│ └── agents/ supervisor + 4 workers (three-layer: prompt, transform, validate)
├── infra/ Bicep + azd (Foundry GA + content filter, Search, KV, ACA, App Insights)
├── evals/
│ ├── quality/ golden cases + CI gates from accelerator.yaml.acceptance
│ └── redteam/ XPIA + jailbreak + brief-specific RAI cases
├── patterns/
│ ├── single-agent/ variant: when orchestration isn't needed
│ ├── chat-with-actioning/ variant: conversational front-end with tools
│ └── sales-research-frontend/ reference UI starter (React + Vite + TS) for /research/stream
├── docs/
│ ├── getting-started/ orientation + setup-and-prereqs (authoritative)
│ ├── partner-playbook.md end-to-end partner motion (7 stages)
│ ├── discovery/ discovery kit (5 artifacts + how-to-use sequencing guide)
│ ├── references/ reference scenarios (customer service, RFP response)
│ ├── agent-specs/ per-agent Foundry bootstrap specs (flagship + candidates)
│ ├── foundry-tool-catalog.md when-to-use matrix for Foundry Agent Service tools
│ ├── customer-runbook.md day-2 ops for the customer team
│ ├── enablement/
│ │ └── hands-on-lab.md partner-team self-paced first-deployment walkthrough (7 labs)
│ ├── patterns/ architecture · WAF · RAI · Azure AI Landing Zone
│ └── version-matrix.md known-good SDK pins (weekly CI validates against latest)
├── .github/
│ ├── copilot-instructions.md hard rules: Agent Framework, MI, HITL, evals, RAI
│ ├── chatmodes/ discover-scenario, scaffold-from-brief, delivery-guide, add-*, switch-to-variant
│ └── workflows/ lint, evals, deploy, version-matrix (weekly pinned-latest)
├── AGENTS.md IDE-agnostic mirror of copilot-instructions (Cursor/Claude/Codex)
└── scripts/
├── accelerator-lint.py ~30 deterministic policy checks (local + CI), AST-only
└── scaffold-scenario.py materialize a new scenario skeleton (CLI behind /scaffold-from-brief)
| Without the accelerator | With the accelerator |
|---|---|
| Partner re-invents auth, telemetry, HITL, evals, RAI posture every engagement | Ships as partner-owned source in src/accelerator_baseline/; used from day one |
| Discovery notes disconnected from code | Solution Brief drives scaffolding, evals, manifest, dashboards |
| "Should we use single-agent or supervisor?" → guesswork | Flagship + two variants + four reference scenarios; pick-and-scaffold |
| Compliance & WAF done at the end (if at all) | Enforced from commit 1 via copilot-instructions.md + CI lint + IaC content filters |
| ROI promises are slides | KPIs declared in accelerator.yaml.kpis[]; partners wire a telemetry event per KPI in the scenario code, then monitor in App Insights + the shipped workbook template (infra/dashboards/roi-kpis.json) |
- customer-service-actioning/ — multi-agent service assistant that looks up orders, issues refunds/credits via HITL, updates CRM. Deflection + AHT ROI.
- rfp-response/ — multi-specialist (pricing · legal · tech · security) aggregator that drafts proposal responses. Response time days → hours; win rate lift.
Flagship itself (sales research & outreach) is fully runnable under src/scenarios/sales_research/ — loaded at startup via the top-level scenario: block in accelerator.yaml. Add a sibling scenario with python scripts/scaffold-scenario.py <id>; the framework mounts it the same way the flagship is mounted.
- Zero Trust posture analysis — chat-based, file-upload (CSV/Excel) assessment with multi-turn iteration. Fits a different solution shape than the flagship (conversational + artifact ingest). Tracked in
docs/agent-specs/README.md; promote todocs/references/zero-trust/when a customer engagement motivates it.
- Not a runtime platform. No services Microsoft operates for partners.
- Not a cryptographic attestation or governance gate. Consistency is enforced by CI lint + pinned SDK + starter defaults + Copilot shaping — not by Microsoft blocking partners at deploy time.
- Not a DSL.
accelerator.yamlis ~12 fields of plain YAML. Nospec.agent.yaml. - Not IDE-locked. Copilot-first; AGENTS.md mirrors the rules for Cursor, Claude Code, Codex CLI.
- GitHub Issues are the intake for scenario requests, bug reports, pattern suggestions.
- Monthly triage; quarterly blessed-pattern promotions (criteria in
CONTRIBUTING.md). - Version matrix is maintained weekly; deprecation policy is N-1 minor.
See SECURITY.md for vulnerability reporting and SUPPORT.md for channels.