Skip to content
Nubent

/DevOps and platform

AI agents for DevOps that run your AWS with human approval

This is how Nubent works internally, and we bring it to your team: AI agents that propose Terraform changes by pull request, triage alerts and incidents, review costs and write runbooks and postmortems. Engineers design, review and approve every change before it reaches production.

/Diagnosis

Signs it is time

The team lives in firefighting mode

Alerts pile up, triage depends on two people and planned platform work slips week after week.

Terraform grows faster than reviews

Infrastructure changes wait days for review, and module and provider upgrades never get done.

Runbooks and postmortems nobody writes

Each incident leaves little documentation behind, and the know-how for fixing problems lives in a few people's heads.

AWS costs nobody reviews in time

Orphaned resources and oversized instances only show up when the bill arrives, because nobody has time to check Cost Explorer every week.

/What we deliver

Concrete deliverables, in your repository

  • Infrastructure change agent

    An agent that proposes Terraform changes by pull request, with terraform plan, security scanning and an impact summary for the reviewer.

  • AI-assisted alert triage

    An agent that picks up CloudWatch alerts, correlates logs, metrics and recent deployments, and posts a summary with hypotheses in the team channel.

  • Recurring cost review

    A report built on Cost Explorer and Compute Optimizer covering waste and savings opportunities, with proposed changes ready for review.

  • Runbooks and postmortems

    Drafts generated from the incident timeline and change history, then reviewed and published by the team.

  • Guardrails and audit trail

    Read-only permissions by default, protected branches, mandatory human approval and a log of every agent action.

/How we work

Our process

  1. Operations assessment

    With read-only access, we review your accounts, repositories, alerts and recent incidents, and pick which repeatable tasks to hand off first.

  2. Agent design and boundaries

    We define what each agent does, with which permissions, which MCP tools it uses and what always requires human approval.

  3. Implementation by pull request

    We build the agents and their infrastructure in your repository, test them on real cases and turn them on gradually.

  4. Ongoing operations

    We track response times, proposal quality and team workload, and refine the agents with every lesson learned.

What doing DevOps with AI actually means

Doing DevOps with AI starts by identifying repeatable operational work, such as reviewing Terraform plans, investigating an alert, comparing costs or drafting a postmortem, and assigning it to agents with scoped permissions. No agent gets administrator access, and people still decide what gets applied and when.

At Nubent we work this way every day. Agents generate and review Terraform, prepare pipelines, analyze costs and document incidents, while engineers design the architecture, review every proposal and approve what reaches production. What we offer is bringing that same operating model to your team, adapted to your accounts and processes.

AI agents for infrastructure as code

Terraform is the most direct use case. An agent can prepare a change requested in an issue, bump provider and module versions, or fix security findings. In every case the output is a pull request with the Terraform plan, results from tools like Checkov and an impact summary written for the reviewer.

That pull request follows the same rules as any engineer's: protected branches, required review and a CI pipeline. The agent never applies changes to production on its own. That clears the infrastructure backlog faster without losing traceability or control over what changes in your AWS accounts.

Alert triage, incidents and postmortems

When an alert fires, the first stretch of work is usually mechanical: check the metrics, search the logs for errors, look at what was deployed in the last few hours. An agent with read-only access to CloudWatch, CloudTrail and deployment history can do that legwork and leave a summary with ranked hypotheses before the on-call engineer even opens the console.

After the incident, the same agent assembles the timeline and a draft postmortem from messages, alerts and applied changes. The team corrects it, adds the underlying causes and defines action items. Over time, those documents feed runbooks that make the next triage better.

Cost and security reviews with agents

Reviewing AWS costs every week is something everyone knows they should do and almost nobody does. An agent can query Cost Explorer and Compute Optimizer, spot idle resources, oversizing or unexpected growth, and propose specific changes by pull request, with estimated savings based on AWS data.

The same applies to security findings. An agent can read AWS Security Hub, group findings by root cause, prioritize the ones that matter and propose the fix in Terraform. Final approval still belongs to a person, but the team stops starting every review from scratch.

Guardrails: permissions, human approval and auditing

Trust in agents is built on clear limits. By default they work with read-only roles in AWS and can only write to their own branches in the repository. Any change to infrastructure, configuration or cost goes through a pull request and needs an engineer's approval.

Every agent action is logged: what it looked up, what it proposed and who approved it. We use least-privilege MCP servers to connect agents to your tools, and we regularly review their output to adjust instructions and scope. If an agent isn't pulling its weight, it gets switched off or redesigned.

/FAQ

DevOps with AI agents: frequently asked questions

Another question? Write to hello@nubent.com and an engineer will answer.

What DevOps tasks can an AI agent handle?

The ones that are repeatable and verifiable: proposing Terraform changes, updating infrastructure dependencies, reviewing plans, first-pass alert triage, cost analysis, grouping security findings and drafting runbooks and postmortems. Architecture decisions, change approval and the final incident response stay with engineers.

Is it safe to let an AI agent touch infrastructure?

It is, as long as the agent never applies changes on its own. Our agents work with read-only permissions in AWS and propose changes by pull request, which go through CI and mandatory human review before they are applied. Every action is logged, and permissions only expand when there is evidence the agent is working well.

What tools are used to build DevOps agents?

We work with Anthropic's Claude models through Amazon Bedrock, with Claude Code and the Claude Agent SDK to build the agents, and with MCP servers to connect them to AWS, the repository and your observability and ticketing tools. Infrastructure is defined in Terraform, and workflows run in GitHub Actions or on AWS services.

Do AI agents replace the DevOps or SRE team?

No. Agents take on the repeatable work, such as gathering data, preparing changes and writing drafts, so the team can spend more time on design, reliability and platform improvements. There is always a person who reviews and approves, and the judgment about what changes in production stays human.

How do I start using AI agents in my operations?

We start with a read-only assessment of your accounts, repositories and recent incidents, and pick one or two high-volume, low-risk tasks, such as first-pass alert triage or Terraform plan review. We measure the results over a trial period and, if it works, add more tasks.

Next step

Let's talk about your DevOps with AI agents project

30 minutes with an engineer. Concrete next steps, whether or not you work with us.