SSLZ Agent documentation · Launch the agent · Classic direct-operator site
Startup Agent Flow
Implemented and planned agent-assisted Azure account and SSLZ setup
Status
Phases 0-6 have delivered contract, planning, or narrowly approval-gated implementation. Phase 7 has delivered execution-disabled cool-foundation and Container Apps planning artifacts, but no secondary-region executor. The PostgreSQL migration, image/CI, dual-cloud connectivity, and program-lineage surfaces are also delivered as planning and validation contracts without execution authority. See the authoritative implementation and evidence matrix.
Start the local conversation from Use SSLZ Agent. Initial planning requires Git and GitHub Copilot CLI authenticated with a GitHub account that has Copilot access, without Azure CLI or Azure login. Install Node.js when running local planners or tests. Live account inspection and preview require additional tools and separate approval. The website does not host the agent or supply an Azure identity.
The integrated validation command in this checkout is:
node scripts/validate-greenfield-journey.mjs
It runs the integrated synthetic founder journey through repository contracts with deterministic mocks, including
fail-closed negative journeys, a separately signed synthetic observed AKS acceptance, and an execution-disabled
cross-program lineage envelope. It performs no Azure writes or live-tenant reads. It requires Node.js and the local
Bicep CLI installed by az bicep install, but no npm install or project dependency restore. Tagged releases expose only
the capabilities documented in their tag; direct Bicep/Terraform usage remains the baseline deployment workflow and is never an implicit agent run.
Run this check in an isolated checkout if an earlier report must be preserved. It deletes and recreates
.sslz/generated/greenfield-journey, and creates and removes the process-specific replay store
.sslz/remediation-state/greenfield-${process.pid}. The agent needs explicit approval for both local write targets.
The fixture’s fixed clock and synthetic approvals do not establish current tenant readiness or authorization.
Acceptance-gap context
The one-subscription and billing-benefit scenario was observed while the existing SSLZ baseline was exercised through a generic agent interaction. That interaction did not intentionally invoke the startup preflight, workload, regional, readiness, IaC, or signed-approval commands. It is therefore an acceptance-gap hardening input for the agent-aware flow and no evidence that those commands regressed.
Before this hardening, the agent-aware flow already had deterministic checks, catalog-driven blockers, current readiness evidence, canonical plan digests, immutable deployment manifests, signed single-use approvals, replay resistance, and live tenant/subscription validation. Only a few gaps remained: Phase 1 required a separate prod/nonprod pair, did not publish a versioned subscription/billing topology decision, and could not bind an external benefit-association confirmation to that exact decision. The topology contract and bindings below close only those gaps and do not replace or duplicate the existing workload, regional, IaC, or deployment controls.
Goal
Help an early-stage startup move from redeemed Azure credits to a reviewed SSLZ deployment plan without requiring the founder to understand Azure account, identity, quota, or regional capacity details first.
The agent guides the founder through:
Account discovery
-> account checks
-> workload discovery
-> region and capacity checks
-> deployment plan
-> founder approval
-> SSLZ deployment
-> post-deployment validation
SSLZ remains the declarative deployment engine. The agent collects context, runs checks, explains blockers, and selects an SSLZ configuration. It must not hide or replace Bicep and Terraform plans.
Design principles
- Start small. Target pre-seed to Series A startups with one primary workload, 5-15 engineers, and no dedicated platform team.
- Use startup defaults. Prefer low-cost managed services and the simplest architecture that satisfies the workload.
- Plan before changing Azure. Show the subscriptions, regions, services, estimated platform costs, and intended changes before deployment.
- Require approval. Never deploy infrastructure, assign privileged roles, or enable paid Defender plans without explicit approval.
- Fail with a next action. Distinguish configuration, permission, quota, capacity, billing, entitlement, and service errors. Provide the official guidance or support route for each blocker.
- Keep existing SSLZ behavior stable. Agent support must be additive and backward compatible.
- Know when to graduate. Recommend full Azure Landing Zones when the workload exceeds the boundaries in the Graduation Guide.
Official sources
| Area | Source |
|---|---|
| Startup account setup | Properly Setting Up Your Azure Account |
| Landing-zone design | Cloud Adoption Framework: Azure landing zones |
| Workload quality | Azure Well-Architected Framework |
| Regional design | Azure reliability guidance |
| Available and paired regions | List of Azure regions |
| Infrastructure as code | Bicep best practices and Terraform style conventions |
The implementation should link each check to a stable official source. Blog posts can provide additional explanation but must not be the normative source for account, security, reliability, or service behavior.
Phase 1: Account discovery and readiness
The account flow follows the Microsoft for Startups setup guidance. The agent checks what Azure exposes, but the founder remains responsible for identity proof, credit redemption, and support requests.
| Check | Agent action | Automatic change? | Block deployment? |
|---|---|---|---|
| Azure authentication | Identify signed-in account, tenant, and active subscription | No | Yes |
| Startup subscription topology | Inventory enabled subscriptions and evaluate an explicit one-subscription or prod/nonprod mapping | No | Yes |
| Credit context | Report visible billing profile and credit context when permissions allow | No | Yes if unresolved |
| Company tenant | Confirm tenant identity and collect the intended company domain | No | Yes |
| Secondary admin | Confirm a second administrator exists | No | Yes |
| Custom domain | Confirm the company domain is verified and primary | No | Yes |
| Subscription selection | Confirm the exact subscription ID for each environment; both may be the same only in validated one-subscription mode | No | Yes |
| Tenant consistency | Confirm both subscriptions belong to the intended tenant | No | Yes |
| Effective permissions | Check the roles required by the selected SSLZ modules | No | Yes |
| Resource providers | Report missing registrations and propose registration commands | With approval | Yes |
| Policy access | Confirm the identity can read and deploy required assignments | No | Yes when policies are selected |
Human-only and support actions
The read-only topology decision distinguishes a validated one-subscription startup, an exact prod/nonprod pair, missing targets, tenant mismatch, unavailable billing evidence, unknown benefit association, benefits observed elsewhere, and an ambiguous multi-subscription inventory. A single visible subscription can be safely mapped to both environments only when the user selects it explicitly. If multiple subscriptions are visible, the user must select an explicit prod/nonprod mapping.
Billing visibility does not prove that startup credits apply. A current Microsoft support confirmation must bind the exact topology decision ID and digest through readiness evidence. Azure Billing Support handles billing account/profile visibility. Microsoft for Startups Program Support handles credit activation, entitlement, and benefit-association questions. Only an opaque support reference is retained.
The agent must not attempt to:
- redeem startup credits;
- create a subscription;
- transfer an entitlement to another account;
- move credits or benefits between subscriptions;
- associate subscriptions with another billing profile;
- change billing associations;
- create or verify a company DNS domain without approval;
- create a Global Administrator or subscription Owner assignment without approval;
- infer that credits apply to a subscription when billing evidence is unavailable.
The agent also cannot prove benefit applicability from partial metadata, complete a support case, or make an ambiguous multi-subscription choice on the user’s behalf. Those conditions remain blocked.
Preflight derives provider checks from repeatable --profile selections. The default remains container-apps for
backward compatibility; an AKS plan explicitly supplies --profile aks, which adds
Microsoft.ContainerService to the inspected namespaces without blocking non-AKS journeys.
Phase 2: Workload discovery
Ask only questions that change the architecture:
- Is this the startup’s first production workload?
- Is the application HTTP/event driven, or does it require Kubernetes APIs and cluster control?
- Does it require PostgreSQL, another managed database, or no database?
- Does it use Microsoft Foundry models?
- Does it require customer-managed GPU compute?
- What data-residency boundaries apply?
- What outage duration and data loss can the business tolerate?
- What monthly platform budget should the plan stay within?
Initial profiles
| Profile | Default | Use when |
|---|---|---|
| Container Apps | Preferred | HTTP and event-driven apps without Kubernetes operations |
| AKS | Exception | Kubernetes APIs, custom scheduling, specialized networking, or GPU nodes |
| Foundry managed models | Preferred for inference | A managed model meets the workload and has quota and capacity |
| Customer-managed GPU | Exception | Custom model serving or training requires direct GPU control |
| PostgreSQL | Preferred relational database | Relational data with PostgreSQL compatibility |
The agent must explain why it selected a profile and what requirement would justify a different one.
See Startup Workload Profiles for the complete selection rules.
Phase 3: Region and capacity planning
Region selection is workload-specific. A region pair alone does not guarantee that every required service, model, SKU, or availability-zone feature exists in both regions.
The implemented scripts/startup-regional-plan.sh command evaluates supplied, timestamped evidence without calling
Azure. It reports quota and point-in-time capacity separately and never treats observed capacity as reserved.
For each candidate region, check:
- subscription access to the region;
- allowed-location policies;
- required service availability;
- Foundry model and deployment-type availability;
- compute and GPU SKU eligibility;
- quota headroom;
- availability-zone support where the workload requires it;
- data-residency constraints;
- estimated cost differences.
Hot/Cool startup topology
The Hot/Cool option is an opt-in conceptual target and not the default for every startup. The table describes intended topology and does not show deployment evidence or execution support.
| Layer | Primary region (Hot) | Secondary region (Cool) |
|---|---|---|
| SSLZ baseline | Deployable through the approved Phase 6 primary path | Planning representation only |
| Networking | Deployed | Deployed with non-overlapping address space |
| Observability | Active | Ready to receive regional telemetry |
| Application compute | Active and scaled normally | Minimum viable scale or deployment-ready, based on RTO |
| Data | Primary | Service-supported geo-replica, geo-backup, or restore plan based on RPO |
| Foundry models | Active deployment | Validated capacity or a documented alternative model/region |
| Ingress | Active endpoint | Registered backend with health checks when automated failover is required |
The plan must state that deploying a landing-zone baseline in two regions does not provide workload failover by itself. Data replication, model availability, secrets, DNS or global ingress, and recovery procedures require service-specific design.
When not to use Hot/Cool
Stay single-region when:
- the startup can restore within its required recovery window;
- the service does not support a useful secondary-region design;
- the additional operational work exceeds the business impact of an outage;
- the startup has not defined an owner for failover tests.
Use a secondary region as a capacity fallback only after validating that every required service and model is available there.
See Hot/Cool Regional Topology for entry criteria, service-specific recovery, cost controls, and testing requirements.
The regional planner marks only current single-region-ready output as executable readiness. A cool-infrastructure
or warm-workload recommendation remains review-only and performs no Azure operation. Separate Phase 7 planners can
generate execution-disabled provider representations, but Phase 6 will not preview or apply them.
Phase 4: Plan and approval
Before deployment, present:
- detected tenant and account context;
- exact target subscription for each environment;
- topology decision ID, digest, freshness, and benefit-association status;
- selected workload profile and rationale;
- primary and secondary regions;
- services, SKUs, and paid security plans;
- quota and capacity findings;
- estimated recurring platform cost;
- Bicep what-if or Terraform plan;
- manual and support actions still required;
- rollback and teardown approach.
When Defender for Servers is selected, Phase 4 also presents a deterministic Log Analytics workspace decision. It blocks unless the exact effective region is allowed by current policy, matches the selected primary region for new placement, is supported by the service and data-residency plan, or is backed by current scope-bound central/shared workspace evidence. Azure default workspace placement is never treated as evidence.
Approval applies to a specific plan. Any change to the topology decision, subscription mapping, tenant, region, paid plan, privileged role, or resource scope invalidates the approval and requires a new plan.
Phase 5: Approved provider remediation
The implemented first write-capable command is limited to one resource-provider registration already present in the
reviewed plan. The provider namespace must be required by the selected workload profile. Apply requires a separate,
unexpired, single-use approval artifact bound to the plan digest and every action field, rechecks the exact tenant and
subscription, runs one argument-array Azure CLI command, and verifies Registered. It performs no deployment,
feature registration, policy, role, billing, entitlement, subscription, or domain change.
See Approved Provider Remediation.
Phase 6: Deployment and validation
Phase 6 is implemented by scripts/startup-deployment-integration.sh. It consumes an approved Phase 4 artifact, emits
an immutable provider-specific manifest through zero-write preview, and requires a trusted Ed25519-signed approval for
apply. It supports only the primary single-region-ready platform baseline.
The agent uses the existing SSLZ Bicep or Terraform path. It must:
- preserve idempotency;
- use workload identity federation for CI/CD;
- avoid secrets in source, parameters, logs, and agent output;
- deploy in incremental mode for Bicep;
- retain Terraform state in an approved remote backend for team use;
- record the plan, approval, deployment IDs, and validation result;
- validate the deployed state instead of treating a successful command as success.
Post-deployment platform validation includes:
- expected resource groups and regional resources;
- policy assignments and compliance state;
- Log Analytics and Activity Log forwarding;
- selected Defender plans, their cost, and the exact approved workspace association;
- budgets and alert recipients;
- explicit
workloadDeploymentAllowed: falseuntil every baseline check passes.
Application health, database connectivity, secondary-region deployment, and Hot/Cool recovery are workload or Phase 7 work and are not performed by Phase 6.
See Approved Deployment Integration.
Cross-program planning lineage
The canonical journey references a separate v1 program-lineage envelope instead of copying migration identities into the signed Phase 5/6 deployment approval. The existing approval scope stays unchanged, and the envelope binds one exact sequence of real planner outputs:
- PostgreSQL migration assessment and planning;
- PostgreSQL rehearsal planning;
- PostgreSQL execution-contract planning;
- container image and CI/CD migration planning;
- dual-cloud connectivity, DNS, identity, and egress planning.
Each stage binds the exact artifact digest and predecessor stage digest. The final program identity and envelope digests therefore change when any upstream artifact, target, environment, lineage, order, or cross-program binding changes. Duplicate, omitted, out-of-order, stale, replayed, mismatched, or substituted artifacts fail closed. The fixtures invoke the repository planner modules with sanitized synthetic evidence; they do not assert against planner source text.
The envelope does not authorize execution. Every stage sets executionEnabled, executionEligible, and
executionAllowed to false and names a distinct future approval. The existing baseline approval remains limited to
greenfield-platform-deployment-only and does not authorize migration, database writes, image promotion, connectivity,
DNS, identity, egress, cutover, rollback, or failback. Its report separately identifies baseline greenfield deployment
readiness and non-executable migration/dual-cloud planning readiness, including whether evidence is synthetic or live.
The canonical greenfield journey report is v2 and requires this envelope identity and readiness separation. A v1 report is intentionally rejected because its lineage is incomplete. See Program Lineage Envelope.
Agent result contracts
The versioned machine-readable result schemas are implemented. startup-preflight produces the inspect result;
scope-limited commands produce the separate provider-remediation, deployment, readiness, regional-attempt,
greenfield-report, and program-lineage contracts linked from the SSLZ Agent overview. The complete
shared preflight semantics, including reserved modes without a generic producer, are in the
Preflight Result Contract. Use the checked-in
ready-container-apps.json and
blocked-billing.json specimens rather than copying an abbreviated contract.
Do not include access tokens, secrets, full billing records, personal email addresses, or other unnecessary personal data in this output.
Delivered baseline criteria
The delivered contract and agent-gated planner baseline can:
- inspect and plan without changing Azure;
- identify the active account, tenant, and explicit target subscriptions;
- detect tenant mismatch and missing required roles;
- report missing providers without registering them automatically;
- classify billing and entitlement uncertainty as a support action;
- ask the workload questions and select Container Apps or AKS with a reason;
- evaluate a primary and optional secondary region without claiming capacity is guaranteed;
- produce a reviewable deployment plan;
- require explicit, artifact-bound approval before an agent-gated write operation.
Direct operator Bicep and Terraform commands remain a separate, non-agent path. The deploy-bicep.yml and
deploy-terraform.yml workflows require Phase 6 artifacts and invoke the approval-gated deployment integration. The
separate integration-test.yml write path is a manually dispatched disposable-nonproduction test path protected by
integration-nonprod; it is outside the agent approval flow and is not a landing-zone delivery path.
Out of scope or still separately gated
- automatic credit redemption or entitlement transfer;
- unattended privileged role assignments;
- automatic DNS-domain verification;
- guaranteed capacity reservation;
- automatic production failover;
- migration of running workloads between regions;
- replacement of the existing SSLZ Bicep or Terraform deployment paths;
- full Azure Landing Zone architecture.
See the Startup Agent Implementation Plan for the phased pull request, testing, approval, and rollback sequence.