[EPIC] azd ai GA: hosted agents and core Foundry capabilities
- Dominant language
- Go
- Stars
- 569
- Forks
- 364
- Avg merge
- 2d 19h
- Merged PRs (30d)
- 136
Description
## Problem statement
`azd ai` is the umbrella CLI experience across the `azure.ai.*` extensions. Customers need a coherent, supported path across hosted agents and core Foundry capabilities - not isolated commands that work only in one extension at a time.
GA means the required customer jobs are reliable interactively and in CI/CD, use stable contracts and GA platform APIs, produce actionable diagnostics, and have complete documentation and operational support.
## GA scope
| Capability | Required GA jobs |
|---|---|
| Hosted agents | Initialize from samples or existing code; add to an application; run and inspect locally; deploy and invoke supported protocols; update versions and endpoints; manage supported sessions and files; diagnose, monitor, delete, and clean up |
| Projects | Create and intentionally name a greenfield project; select and reuse a brownfield project; manage model, region, quota, and environment configuration; own and reconcile `azure.ai.project` in `azure.yaml` |
| Connections | Create, read, update, delete, and consume supported connection categories and authentication modes; own and reconcile `azure.ai.connection` |
| Toolboxes | Create, compose with connections, version, publish, attach to an agent, and successfully invoke a supported tool |
| Routines | Create, update, inspect, enable, disable, dispatch, review runs, delete, invoke the configured agent action, and reconcile through `azure.yaml` |
| Inspector | Launch against every supported local protocol and fail clearly when the target is unavailable or incompatible |
### Teams and Microsoft 365 bot path
Teams and Microsoft 365 publication is part of the hosted-agent experience. An Activity-protocol agent must support this end-to-end path:
- [ ] Configure the Activity protocol and Azure Bot Service.
- [ ] Produce validated handoff inputs and guidance.
- [ ] Use approved Microsoft 365 tooling for personal sideload or tenant-catalog publication.
- [ ] Verify a successful Teams or Microsoft 365 interaction.
`azd` does not need to upload the application itself.
## Required reference experiences
- [ ] Install or upgrade a compatible `azd` version and the complete GA extension set without breaking an existing project.
- [ ] Create an intentionally named greenfield Foundry project with supported region, model, quota, and environment configuration.
- [ ] Select and reuse a brownfield Foundry project without taking ownership of reused resources.
- [ ] Initialize, deploy, and invoke every supported hosted-agent protocol from both a sample and existing code.
- [ ] Add a hosted agent to a multi-service application and round-trip the stable unified `azure.yaml` representation.
- [ ] Run and inspect supported agents locally.
- [ ] Create and consume supported connections.
- [ ] Create a toolbox, compose it with connections, publish it, attach it to an agent, and invoke a supported tool.
- [ ] Complete the routine lifecycle from creation through dispatch, run inspection, and deletion.
- [ ] Complete the supported non-interactive CI/CD path with secure authentication, validation or preview, deterministic deployment, structured output, environment promotion, and safe redeploy or rollback.
- [ ] Delete owned resources while preserving reused brownfield resources.
## GA success criteria
| Area | GA bar |
|---|---|
| End-to-end reliability | At least 95% success for supported hosted-agent deployment journeys |
| New-user reliability | At least 90% success on the first deployment attempt |
| Reference journey speed | Greenfield `init -> deploy -> successful invoke` completes at p90 in 10 minutes or less |
| Recovery | Median attempts to success is 1; p90 is at most 2 attempts; retry-to-success p90 is at most 30 minutes |
| Cancellation and hangs | Cancellation remains at or below 1%, with no known indefinite waits in supported interactive or non-interactive paths |
| Scenario quality | Required end-to-end scenarios sustain at least 98% aggregate success, with no required scenario below 95% |
| CI/CD | A supported reference path completes secure authentication, validation, deterministic deployment, structured output, environment promotion, and safe recovery |
| Contract stability | Command names, exit behavior, JSON schemas, unified `azure.yaml`, API versions, extension versions, migration, and deprecation behavior are stable and tested |
| Quality and security | No unresolved critical reliability, security, privacy, data-loss, resource-leak, or destructive-cleanup defect affects a required experience |
| Documentation and support | Getting-started, command reference, troubleshooting, diagnostics, runbooks, escalation, rollback, and support handoff are ready |
| Customer validation | External customers successfully complete the relevant GA experiences |
## Execution workstreams
- [ ] Freeze contracts, GA API versions, supported variants, command naming, and per-extension ownership.
- [ ] Complete all required interactive customer experiences.
- [ ] Complete the supported end-to-end CI/CD path.
- [ ] Build continuous scenario and reliability evidence.
- [ ] Complete documentation, diagnostics, operations, support, and rollback readiness.
- [ ] Complete external customer validation.
Contributor guide
Assessment
This issue has not been assessed yet.