Forum Discussion

pavanaddepalli's avatar
pavanaddepalli
Occasional Reader
Nov 19, 2026

Azure-SRE-Agent-AI-Native-SDLC-Blog

Beyond incident triage: Seven practices for an AI-native SDLC with Azure SRE Agent

Explore a practical approach to AI-assisted operations using reusable prompts, bounded autonomy, traceable actions, measurable quality, operational feedback, and cost governance.

AI-assisted operations should be evaluated by more than the quality of an answer. The important questions are whether the answer is grounded in evidence, whether actions stay within authorized boundaries, and whether the outcome can be verified.

Azure SRE Agent brings together context from Azure resources, observability tools, incident platforms, and source repositories. It can investigate issues and propose or execute mitigations according to the permissions, run modes, and policies you configure. [1]

This article proposes seven practices for connecting those capabilities to an AI-native software development lifecycle (SDLC). Treat them as an adoption framework, not as a claim that every control is enabled by default.

1. Make prompts reusable and evidence-focused

Start with a small, version-controlled prompt catalog for recurring tasks: incident investigation, test generation, code review, and documentation.

For each template, specify the task, resource scope, time range, expected output, and behavior when evidence is missing.

Here is an illustrative investigation prompt:

Investigate failures in {PIPELINE_NAME} between
{START_TIME_UTC} and {END_TIME_UTC}.

Scope: {RESOURCE_GROUP} and {REPOSITORY}.

Use read-only operations. Correlate failed runs, dependency
health, and recent changes.

Return:
- Observed symptoms and supporting sources
- Ranked hypotheses and evidence gaps
- Recommended next checks
- Proposed mitigation, risk, and rollback approach

Do not change resources or create a pull request.
If evidence is insufficient, state what is missing.

Use this template as an investigation contract. Evaluate whether it produces relevant evidence and useful next steps before using it in an automated workflow.

2. Turn guardrails into enforceable controls

Write down what the agent may read, what requires approval, and what must remain unavailable. Then map those decisions to identity permissions, tool policies, and workflow controls.

Azure SRE Agent supports tool access policies alongside run modes, hooks, and connector governance. These controls interact, so a prompt instruction alone should not be treated as an authorization boundary. [2]

For an initial rollout, adopt a deliberately narrow policy: read operational signals, draft proposed changes, and leave consequential actions behind a verified approval path.

3. Define bounded autonomy explicitly

Use a risk model to decide how much authority each workflow needs.

Suggested category

Intended behavior

Read-only investigation

Gather evidence within the approved resource scope

Reversible nonproduction action

Execute only under an explicitly configured policy

Production change

Present the change, impact, verification, and rollback plan for approval

Prohibited action

Block execution and escalate to the responsible owner

 

This is a suggested operating model, not an Azure SRE Agent configuration schema.

Test effective permissions rather than relying on labels. Current documentation describes policy precedence and approval behavior that vary with run mode and hooks; for example, an allow policy can permit execution in Review mode. [2]

4. Design for traceability

For your implementation, define a provenance record that connects:

Request -> evidence -> proposed action -> authorization
        -> execution result -> verification -> reviewed change

Record resource scope, timestamps, tool outcomes, approval decisions, and links to resulting artifacts. Capture concise explanations and supporting evidence, not an assumption that all internal model reasoning is available.

Azure Monitor's agent observability capabilities provide a foundation for investigating execution, response quality, failures, and consumption across compatible instrumented agents. Validate which telemetry your particular integration emits. [3]

5. Measure correctness separately from completion

A workflow that finishes is not necessarily a workflow that succeeds.

Use an evaluation scorecard with separately defined measures:

Measure

Suggested evidence

Root cause accuracy

Comparison with the reviewed incident resolution

Query correctness

Successful execution plus relevant results

Change quality

Tests, reviewer findings, and substantive rework

Authorization compliance

Action records matched to required approvals

Resolution time

Explicit start and verified-recovery timestamps

Unsupported claims

Human review against cited evidence

 

Set thresholds from an initial baseline. Publish sample sizes, evaluation periods, and exclusions alongside results.

Avoid combining these into a single score until the normalization rules are clear. Keep authorization violations visible independently rather than allowing other scores to offset them.

6. Close the loop with reviewed operational knowledge

Azure SRE Agent can reference uploaded documentation, saved memories, and learnings from earlier investigations. These capabilities help carry operational context between conversations. [4]

Use a reviewed knowledge record for recurring incident patterns:

Symptom:
Confirmed cause:
Evidence:
Resolution:
Verification:
Conditions where this guidance does not apply:
Owner and review date:

Treat historical knowledge as a starting hypothesis. Require current evidence before applying a previous resolution to a new incident.

Extend the loop to the SDLC: propose a regression test, improve a runbook, or correct a monitoring gap. Review those changes through the same process as other engineering work.

7. Optimize cost without removing necessary evidence

Azure SRE Agent usage is measured in Azure Agent Units (AAUs). Its consumption views help identify expensive conversations, activity types, and usage trends. Token counts are useful diagnostics, but they are not a complete billing estimate. [5]

For your workflows, start with narrow queries, bounded outputs, relevant knowledge, and appropriately scheduled automation. Compare each optimization against the same evaluation cases used to measure quality.

A useful adoption metric is cost per correctly completed task, reported alongside investigation accuracy and resolution time. Lower consumption is not an improvement if it removes evidence needed for a reliable decision.

Start with one measurable workflow

Choose a bounded investigation scenario in a nonproduction environment. Establish its prompt, permissions, evidence requirements, evaluation cases, and consumption baseline.

Expand authority only after the workflow demonstrates useful results and its approval boundaries behave as intended.

The goal is not maximum autonomy. It is an engineering process in which AI-assisted work is grounded, authorized, traceable, and verifiable.

No RepliesBe the first to reply