[Initiative] 🎯 Deliver reliable end-to-end Azure agent workflows

## Problem statement

A successful Azure agent experience requires more than invoking the correct skill. Developers need reliable routing, context transfer, tool selection, confirmations, error recovery, and clear completion across multi-step workflows. Breakdowns between these stages create incomplete or unsafe outcomes even when each individual component works as designed.

## Vision

Developers can express an Azure goal and receive a coherent, safe, end-to-end workflow. GitHub Copilot for Azure selects the right capabilities, preserves relevant context, handles expected failure modes, and verifies the requested outcome before declaring success.

## Who this helps

- **Azure developers and operators** can complete multi-step tasks with fewer manual handoffs.
- **Skill authors** get shared patterns for composition, state, recovery, and verification.
- **Security and reliability teams** get explicit controls around risky operations and failure behavior.
- **Maintainers** can measure workflow success rather than isolated skill invocation.

## Goals (in scope)

- Improve intent routing and capability selection for multi-step Azure tasks
- Define composition and handoff patterns between skills, tools, and agents
- Preserve necessary context while minimizing sensitive or irrelevant data
- Apply explicit confirmation, least privilege, and safe failure behavior to consequential operations
- Add recovery and verification patterns for common Azure failure modes
- Measure end-to-end completion, correctness, and user effort

## Non-goals (out of scope)

- Guaranteeing deterministic model behavior
- Adding broad Azure service coverage without a validated user journey
- Replacing host-client orchestration capabilities
- Hiding errors or returning success-shaped fallbacks when work is incomplete

## Success criteria

- [ ] Priority workflows have documented entry points, steps, trust boundaries, and completion criteria
- [ ] Routing and composition evaluations cover representative positive and negative scenarios
- [ ] Consequential operations consistently use confirmation and least-privilege guidance
- [ ] Expected failures produce actionable recovery paths without losing workflow state
- [ ] End-to-end completion and correctness improve against maintained baselines
- [ ] User-reported workflow failures can be reproduced as tests or evaluations

## Dependencies

- **Copilot orchestration surfaces**: Workflow composition depends on supported host capabilities
- **Azure MCP and service tools**: Reliable execution depends on stable contracts and actionable errors
- **Evaluation infrastructure**: End-to-end quality requires representative datasets and graders
- **Security guidance**: Trust boundaries and confirmation patterns require ongoing review

Start by reviewing the priority workflows and the supported Copilot orchestration surfaces, Azure MCP and service-tool contracts, evaluation infrastructure, and security guidance. Define documented entry points, steps, trust boundaries, and completion criteria for a validated workflow. Done means the applicable success criteria are met, including routing evaluations, safe confirmations, recovery paths, and reproducible workflow tests or evaluations.

Written by the indexing model from the issue text.