[Hosted agents 05] Add reusable hosted-agent infrastructure primitives

[placerda]

Implementation scope

• Repository: Azure/bicep-ptn-aiml-landing-zone
• Base branch: develop
• Prerequisites: [Hosted agents 01] Freeze implementation contracts and release gates #588 and [Hosted agents 03] Add hosted Responses adapters and managed Conversations #598 (completed)

Summary

Add reusable Microsoft Foundry hosted-agent infrastructure primitives to the AI Landing Zone without encoding GPT-RAG product topology or lifecycle policy.

Repository boundary

The AI Landing Zone serves multiple accelerators. It may provision generic hosted-agent capabilities, but it must not know about GPT-RAG chat backends, administrative panels, strategy names, Cosmos ownership, or the decision to remove an application workload.

The following belong exclusively to Azure/GPT-RAG and are out of scope here:

• DEPLOY_ADMINISTRATIVE_PANEL / deployAdminPanel;
• the classic, hosted/no-panel, and hosted/panel product modes;
• conditional removal or retention of the GPT-RAG orchestrator Container App;
• GPT-RAG-specific Cosmos, feedback, history, configuration, naming, or lifecycle behavior.

Required generic capability

• Provide typed, reusable hosted-agent resource inputs for image, immutable version, entrypoint/runtime settings, identity, networking, and optional deployment enablement.
• Provision generic managed identities and least-privilege RBAC required by the hosted-agent resource itself.
• Expose stable generic resource identifiers/endpoints consumed by an accelerator.
• Preserve existing Container Apps and data resources exactly as requested by their existing generic inputs; enabling hosted-agent infrastructure must not implicitly remove or retain another workload.
• Preserve private endpoints, DNS, egress, and ACR public-access controls.
• Support the ADR-approved generic private-build primitives: VNet-connected runner/agent baseline and jump-host fallback. Shared ACR Tasks are not a private-endpoint bypass.
• Default all new capability to disabled so existing accelerators are unchanged.

Acceptance criteria

• No GPT-RAG-specific parameter, resource name, mode table, panel reference, Cosmos policy, or workload-removal condition exists in this repository.
• Bicep and parameter contracts are typed and accelerator-neutral.
• Existing topology renders unchanged when hosted-agent capability is disabled.
• Tests/fixtures prove hosted-agent resources can be added without modifying unrelated Container Apps or data resources.
• Network-isolated paths keep ACR public access disabled.
• Focused and full existing validation pass, with exact commands and residual live deployment gates recorded here.

GPT-RAG composition is owned by #595 in Azure/GPT-RAG.

[placerda]

Implementation status

Implementation is underway in Azure/bicep-ptn-aiml-landing-zone from develop:

• Pull request: feat: add accelerator-neutral Foundry hosted-agent prerequisites bicep-ptn-aiml-landing-zone#121
• Hosted runtime prerequisite: [Hosted agents 03] Add hosted Responses adapters and managed Conversations #598 completed through feat: add hosted Responses adapters and managed Conversations entrypoint gpt-rag-orchestrator#285 at 083cd3f728c695113052d454894bf328e7134bcd

This issue remains the canonical tracking item. Completion evidence will include the merged infrastructure commit, three-mode resource-graph proof, false-default and rollback validation, identity/RBAC and private-network contracts, and residual live deployment gates.

[placerda]

Implementation evidence — architecture boundary satisfied

Draft implementation: Azure/bicep-ptn-aiml-landing-zone#121 at bc8c35df99cf0b2ca146430ccd3ae709e3a8b7ed.

Acceptance result

• All deployAdminPanel behavior, GPT-RAG modes/names, panel/chat/Cosmos policy, workload-removal conditions, and hosted-agent Container App/UAI coupling are absent from the final PR diff.
• deployHostedAgent=false preserves the merge-base resource graph.
• deployHostedAgent=true is additive and does not alter unrelated Container Apps or data resources. It adds no service/workload resources; only two generic gated RBAC entries are contributed through the existing centralized role deployments: deployment principal → Foundry Project Manager at Foundry project scope, and Foundry project identity → Container Registry Repository Reader at selected ACR scope.
• Typed, immutable image/runtime/protocol and Foundry/ACR/network/private-build outputs are exposed for a downstream azure.ai.agent service. The accelerator remains responsible for agent version creation, dedicated identity, endpoint, and runtime AcrPull.

Validation

• Main Bicep and hub fixture compilation passed.
• Template size gate passed at 4,882,930 bytes (4.657 MB).
• Bicep 0.42.1 resource contract passed: 65 symbolic resources, all 63 non-hosted definitions equal the merge-base fixture, with only the two centralized RBAC deployments allowed to receive gated entries.
• Deterministic preflight suite passed 42/42.
• Default offline preflight had 0 failures and one unrelated existing Foundry IQ warning.
• Unsupported hostedAgent.roles lists fail preflight; the sealed type accepts no arbitrary role names.n- The released v2.3.0CHANGELOG section exactly matches the merge base; only this capability is underUnreleased.n- Forbidden-scope scan found no GPT-RAG/admin-panel/product-mode terms outside the test's forbidden-token list.
• bicep-validate run 30455146744 passed.
• Draft public-site companion Azure/AI-Landing-Zones#127 is reduced to 311 additions / 1 deletion; strict MkDocs build and CI passed.
• Two independent final reviews found no high-confidence defects.

Residual live gates

No Azure deployment was performed. Before merge/release, validate or explicitly accept: standard and isolated azd provision in both flag states; downstream prebuilt-digest azd deploy plus invocation/identity/AcrPull smoke test; VNet-reachable private-ACR build/push; the documented June 25, 2026 Foundry project-date gate for private-endpoint-only ACR; and existing-ACR PE/DNS/authentication-as-ARM/RBAC/network reachability.

Release impact is backward-compatible minor; Portal and Terraform need parity review. Both PRs remain draft. Nothing was merged or released.

[placerda]

Completed and merged in Azure/bicep-ptn-aiml-landing-zone#121.

Final evidence:

• Reviewed head: bc8c35df99cf0b2ca146430ccd3ae709e3a8b7ed
• Merge commit on develop: df9c4be20a9f375a15681bafd0a8b2c89ba1256d
• Boundary: accelerator-neutral Foundry hosted-agent prerequisites only; no deployAdminPanel, GPT-RAG mode policy, GPT-RAG Cosmos/history/panel ownership, CHAT_BACKEND, or Container Apps suppression
• Behavior: opt-in and additive; disabled symbolic resource graph remains unchanged; outputs are neutral when disabled
• Security: Foundry Project Manager is scoped to the project and Container Registry Repository Reader is scoped to the selected registry
• Validation: Bicep compilation, deterministic preflight fixtures, forbidden-token/graph-parity contract tests, and the size gate passed
• CI: bicep-validate run 30455146744 and CLA passed at the exact reviewed head
• Independent final review found no merge blockers

Residual live-Azure proof remains intentionally assigned to #597: disposable isolated deployment, private build/push, downstream digest deployment, per-agent identity and AcrPull smoke tests, and rollback validation. No component or umbrella release was published.

[placerda]

Closing as the generic AI Landing Zone primitives are independently reviewed, validated, and merged to develop. GPT-RAG-specific composition remains exclusively in #595.