AI agents for DevOps: what to hand off and what to keep

What an AI agent for DevOps does

On a small team, an AI agent for DevOps is mostly a reader. Point it at your CI and your cloud bill, and it decides what to dig into and tells you what changed. A script only runs the steps you wrote.

If nobody on your team is on call, you probably don't need an AI SRE agent yet. They pull the alert, recent deploys and logs together for whoever got paged. Try the weekly work below first.

Go in a little skeptical. In Stack Overflow's 2025 survey, 46% of developers said they distrust the accuracy of AI tools, and the 2025 DORA report tied AI use to higher delivery throughput but lower stability. That's why everything below starts read-only.

Our test for what to hand off: the job comes up every week, and you can check the answer in a few minutes. That's close to what Google's SRE book calls toil: manual, repetitive work that grows with your service.

What to hand off first

Start with jobs where the agent reads and reports, and you can check its evidence.

For CI, copy the pattern in Harness's AutoFix docs. When a pipeline step fails, the agent reads the failed run and the repository changes, then opens a pull request for a person to review. Watch for one thing. A 'fix' that weakens a test so CI goes green is worse than no fix, so read every test change it makes. No CI basics yet? Start with DevOps for startups.

For alerts, the part worth handing over is ranking the pile. GitHub already lets you assign a Dependabot alert to an agent, which opens a draft pull request with a proposed fix. Security alert triage shows the ranking done by hand, and dependency breaking changes covers upgrades.

For cost, leave AWS's budget and anomaly alerts on and have the agent run the weekly comparison from AWS cost monitoring.

For deploys, watch your change fail rate, which DORA defines as the share of deployments that need immediate intervention after they ship.

What a person keeps

Our rule: if it's hard to undo or it spends money, a person presses the button.

OWASP calls the risk excessive agency: an agent with more access or freedom than its job needs does damage when it misreads its input or gets tricked. Its answer is ours too: a person approves anything high-impact. GitHub's Copilot cloud agent works that way. It can push to one branch only, and it can't approve or merge a pull request (GitHub docs).

Guardrails to ask for

Ask any vendor for these five before you connect anything.

Read-only access

Start read-only. On AWS, that's the ViewOnlyAccess and SecurityAudit managed policies. Neither covers Cost Explorer, so the cost job also needs a few read-only billing permissions such as ce:GetCostAndUsage. On GitHub, add Contents and Pull requests write only when you want draft pull requests, and turn on branch protection first. Then read the permissions in the AWS and GitHub consoles yourself.

Pull requests as the only way in

Turn on branch protection for main before an agent gets any write access. Branch protection can require a pull request, a review and passing checks before anything lands (GitHub docs).

Approval for anything sensitive

You should be able to say no to any action before it runs. MCP's spec keeps a person in the loop: you can always refuse a tool call, and sensitive ones should ask first. Azure SRE Agent lets an admin set each tool to allow, ask or deny (Microsoft Learn).

A record of what it did

Ask the vendor where you'd read a log of every action the agent took and who approved it. The MCP specification calls for logging tool use for audit.

A stop switch

You should be able to pause one agent or everything at once and keep what's saved. Test the pause before you rely on it. For one example, see how pause works in Since, which we make.

What an AI DevOps agent costs

Ask what one run costs before you switch it on. Vendors meter it differently:

• GitHub bills the cloud agent in AI credits by the tokens a session uses, at $0.01 a credit, so a long session costs more than a short one (GitHub docs).
• Datadog works out to about $6.50 to $8.45 per autonomous investigation, by our arithmetic: about 6.5 credits each, at $1 to $1.30 a credit (Datadog pricing).
• Microsoft charges for Azure SRE Agent every hour it runs, busy or not, plus tokens, and a monthly cap doesn't stop the hourly part (Microsoft Learn).

Our suggestion: set a monthly cap or alert on day one, if the vendor offers one, and read the first week's usage before you give the agent more to cover.

Trial an AI agent in a week

A week is enough for a first look. The schedule below is ours. Before an agent acts on its own, Microsoft advises watching what it recommends for two to four weeks.

1. Day 1. Give it read-only access and pick three jobs from the hand-off table.
2. Days 2 and 3. Backtest it on your last 5 to 10 CI failures and alerts, and score each answer with the labels below.
3. Days 4 and 5. Run it live while you still do your weekly operations review by hand, then compare.
4. Day 6. Tally the wrong answers and the minutes saved, and note what it cost.
5. Day 7. Pick one: keep it read-only, allow draft pull requests, or stop.

right              cause and next step were correct
partly right       right area, wrong detail
wrong              cause was wrong, and it hedged
confidently wrong  cause was wrong, stated as fact

These labels are our own, loosely based on Traversal's scorecard, which gives a wrong root cause zero credit. Our suggestion: count "confidently wrong" as the worst result, since a confident answer sends you the wrong way and a hedged one makes you check. If the week shows the work needs a person, read when to hire your first DevOps engineer.

Where Grant fits

We make Since, so read this part with that in mind. Grant and his team are AI agents for software operations. Their check-ins read the sources you connect. David can open a draft pull request for a supported fix, and on Studio and Enterprise Vera tests agents you built yourself. Nothing merges or deploys. Here is what each teammate reads, by job:

• CI failures on main: Maya reads GitHub Actions runs and CI results reported to GitHub on the default branch.
• Alerts: Priya reads existing Dependabot and code scanning alerts and AWS Security Hub findings. On paid plans, David sends supported fixes for upstream changes that affect your code as draft pull requests. On Free you can ask for one on demand, within a monthly limit.
• Cost: Liz reads Cost Explorer by service, and on paid plans AWS's forecast and anomalies. On Studio and Enterprise she also reads OpenAI and Anthropic spend from an admin key the owner adds.
• Deploy checks: on Pro and up, Noah reads CloudWatch metrics for the EC2, Lambda and RDS resources you select. On Studio and Enterprise he compares errors and load after a production deploy, and he may recommend a rollback but never does one.
• The weekly summary: Harper writes the Monday ops brief on Pro and up.

Since's own AI work is counted in credits, and every plan has an allowance: 1,000 a month on Free, 25,000 on Pro and 80,000 on Studio. Free stops there. Pro and Studio carry on past it, and the extra credits go on a separate monthly invoice (pricing).

Want a hard cap? On a paid plan, run Since on your own Anthropic or OpenAI key and set the limit with your provider. That work uses no credits (BYOK vs hosted AI).

On paid plans you can also ask Grant over MCP at https://api.since.dev/mcp, and our MCP guide covers the setup. An owner or admin approves the projects and actions for each connection, and no Grant tool changes your cloud, merges or deploys (MCP docs).

Keep your own alerts for outages. Grant isn't a pager. Scheduled check-ins run daily on Pro and every 6 hours on Studio, and they also start on GitHub events (plans).

Questions

What is an AI agent for DevOps?

It reads your CI, alerts, cloud bill and deploys, decides what to look at next and uses tools to get it, then reports or drafts a change. A script only runs fixed steps. A person still merges, deploys and approves access and spending.

What is an AI SRE agent?

An AI SRE agent helps during an incident: it pulls together the alert, recent deploys and logs for whoever got paged. If nobody on your team is on call, you probably don't need one yet, and we'd start with weekly checks. Rootly's post covers what these agents do well and badly.

Can an AI agent fix my CI failures or deploy for me?

An agent can draft a fix as a pull request, the way Harness's AutoFix does. It shouldn't merge or deploy on its own, and branch protection keeps a pull request the only way into main.

Is it safe to give an AI agent access to AWS or production?

Yes, if it starts read-only with AWS's ViewOnlyAccess and SecurityAudit managed policies. Neither covers Cost Explorer, so add a few read-only billing permissions if you want it to read costs. We'd add write access only for one narrow job, with a person approving it.

How much does an AI DevOps agent cost?

Vendors price it differently, so ask what one run costs. Datadog lists an autonomous investigation at about 6.5 credits, which by our arithmetic is 6.50 to 8.45 dollars (Datadog pricing), and GitHub bills its cloud agent in AI credits by the tokens a session uses (GitHub docs). Set a monthly cap or alert either way.

Will AI agents replace DevOps engineers?

We don't think so. An agent can take the reading and drafting, but someone still has to own uptime and on-call. When to hire your first DevOps engineer lists the signs you need one.

Related guides

• How to connect Claude or ChatGPT to an MCP serverWhat an MCP server is, how to add a remote one to Claude, ChatGPT or Cursor, a read-only GitHub example, and OAuth vs API keys.
• Bring your own key (BYOK) vs hosted AI: which to chooseBring your own key (BYOK) means an AI tool runs on your own OpenAI or Anthropic LLM API key. When it beats hosted AI, and how to set it up safely.