AI Tools for DevOps in 2026: 17 Tools Mapped to the Job They Actually Do

Key points

• AI tools for DevOps cover three behaviours that need different levels of trust: assistants that suggest, diagnostics that explain, and agents that act. Most procurement pain comes from treating them as one purchase.
• Of the 17 tools compared here, three run fully inside your own boundary: K8sGPT, HolmesGPT and NudgeBee. Everything else is SaaS, or an agent reporting to a vendor-hosted backend.
• Nine of the 17 do not publish a price. Where they do, the unit ranges from per seat to per gigabyte ingested, and the ingest unit is where budgets break.
• Very few tools execute changes unattended, and in every case it is opt-in behind guardrails. Read-only investigation carries almost no risk and delivers most of the early value.
• This guide is ordered by the operator's day rather than the pipeline stage, so incident response comes first and coding assistants come last.

What are AI tools for DevOps?

AI tools for DevOps are systems that use machine learning or large language models to do part of the work of building, shipping and running software: writing pipeline and infrastructure code, reviewing changes before they land, investigating incidents, and reducing cloud and cluster spend.

Two very different things get sold under that label. Some write code, and operate on a repository where a bad suggestion costs a review comment. Others run in production, where the same mistake costs an outage. The evaluation questions differ, and so do the deployment constraints.

Most lists organise by pipeline stage: code, build, test, deploy, monitor. This one is ordered by the operator's day instead, starting with the thing that wakes people up.

The short list: best AI DevOps tool by job

• Best for incident investigation you can self-host: NudgeBee
• Best for Kubernetes triage at no cost: K8sGPT
• Best for on-call routing and coordination: PagerDuty
• Best for automated Kubernetes rightsizing: Cast AI
• Best for AI code review on pull requests: CodeRabbit
• Best for writing pipeline and infrastructure code: GitHub Copilot

Comparison table: 17 AI DevOps tools by deployment, data boundary and price

Pricing checked 19 August 2026.

Three things are visible here that no competing page publishes. Most of this category is SaaS-only. Most of it does not publish a price. And "does data leave your boundary" is answerable for every tool but is almost never asked.

How we picked these 17 tools

1. Where it runs. Inside your infrastructure, or as a service you ship telemetry to.
2. What it is allowed to do. Advise, suggest, execute with approval, or execute unattended. And if it executes, where the approval arrives.
3. What it correlates. One data type, or signals across logs, metrics, traces, deployments and infrastructure state.
4. Whether you can read the source and run it yourself.
5. What it costs, and per what unit. Per seat, host, cluster, node, or gigabyte ingested.

Everything here is generally available unless the entry says otherwise. Where a vendor does not publish something, the cell says so rather than carrying an estimate.

Copilot, agent, or diagnostic: the distinction that decides your blast radius

A copilot suggests. It needs read access to a repository, and a bad suggestion costs a review comment.

A diagnostic explains. It needs read access to production, which is a real privilege, but it changes nothing.

An agent acts. It needs credentials that can modify infrastructure, and the question stops being "is the answer good" and becomes "what can this break, who approved it, and can I reverse it".

Most procurement pain in this category comes from treating all three as one purchase.

AI SRE and incident response

Two different products get called incident tooling. One routes and coordinates: who gets paged, how the incident is tracked, what the retrospective says. The other investigates: reading telemetry to work out what broke. Most teams need both, and buying one expecting the other is the most common mistake here. Our comparison of AI SRE tools for reliability engineers covers the landscape by persona, and automated incident response tools covers coordination.

1. PagerDuty: best for alert routing and on-call with AI event intelligence layered on

PagerDuty is the default answer for on-call, and its AIOps layer reduces the noise reaching a human by grouping related alerts into one incident.

• Event correlation and alert grouping
• On-call scheduling, escalation policies and incident workflows
• Automated runbook execution

Use case: forty alerts from one failing dependency arrive as a single incident with one page.

Where it does not fit: correlation groups symptoms, it does not explain them. PagerDuty's Fully Autonomous Responder is early access rather than GA, so shipped behaviour today is execution with human approval. incident.io and Rootly are the modern alternatives on coordination. Rundeck now ships as PagerDuty Process Automation for on-premise and Runbook Automation for cloud.

Deployment: SaaS. Pricing: 14-day trial, not published.

2. Cleric.ai: best for autonomous alert investigation that reports back to on-call

Cleric investigates alerts on its own and reports findings to the on-call engineer, aiming to complete the first twenty minutes of work before a human starts.

• Autonomous investigation triggered by alerts
• Findings delivered into the on-call workflow
• Read-only by default, enforced through access scopes

Use case: an alert fires overnight and the investigation, with evidence gathered, is waiting when someone opens their laptop.

Where it does not fit: it runs as single-tenant SaaS on dedicated infrastructure rather than in your cluster. An optional customer-deployed connector reaches private networks, but the investigation happens on Cleric's side. NeuBird Hawkeye and Resolve AI occupy the same category.

Deployment: single-tenant SaaS. Pricing: not published.

3. NudgeBee: best for self-hosted incident investigation on clusters where data cannot leave the VPC

NudgeBee runs entirely inside your own cluster, investigates an incident to a cited root cause, then applies the fix only after a human approves it. That combination is what fits when telemetry cannot leave your environment at all. It is our product, and this entry is written to the same five criteria as every other.

• Four assistants on one backend: AI-SRE, AI-FinOps, AI-K8sOps and CloudOps, plus an AIOps Automation Builder for custom workflows
• Investigate-not-execute by design: read-only diagnostics run freely, and every create, update or delete is gated behind explicit human approval, enforced in the tool-access layer rather than by prompt alone
• Queries your existing Prometheus, Grafana and Datadog backends in place rather than re-ingesting telemetry
• Knowledge-graph correlation across 61 node types and 37 relationship types, separating root cause from symptom rather than grouping by time window
• Blast-radius and safety scoring before any change, stamped Safe or Review
• Bring your own model across 11 provider routes including Bedrock, Ollama and vLLM, with embeddings able to run on-device
• Outbound-only: the in-cluster agent dials out over a single WebSocket, so no inbound ports and no VPN
• Readable source: you can read the code and run it in your own cluster

Use case: a pod enters CrashLoopBackOff, the AI-SRE assistant correlates cluster state, recent deploys and logs into a cited root cause, then proposes the fix as an approval-gated change with its blast radius attached. Customers report 70% lower MTTR and 30 to 40% lower cloud spend.

Where it does not fit: it is not a code-generation assistant, not a pull request reviewer and not an IDE agent, so it does not compete with tools 13 to 16 on this list. It will not act on infrastructure without a human approving the change, which is deliberate but means it moves at the speed of the reviewer. And it is a newer entrant than the observability incumbents, so the third-party integration library and public case-study record are smaller.

Deployment: fully self-hosted. No telemetry, no phone-home. Pricing: free for up to 2 clusters or cloud accounts, full platform. AI-SRE from $2,000 per month at the 2 to 4 cluster tier, scaling with cluster count. Enterprise custom.

AI for Kubernetes diagnostics and cluster operations

Generic AIOps degrades on Kubernetes for structural reasons. Pods are ephemeral, so the thing that failed may not exist by the time anyone looks. Ownership is ambiguous: a failing pod belongs to a deployment, which belongs to a team, and none of that is in the alert. Label cardinality makes naive correlation expensive. And the common symptoms are not causes. A CrashLoopBackOff tells you a container keeps exiting, not why.

This is also the only part of this keyword cluster growing meaningfully in the US, which suggests the gap is felt.

4. K8sGPT: best free starting point for LLM-assisted Kubernetes diagnostics

K8sGPT scans a cluster, finds problems and explains them in plain language. It is Apache 2.0 and a CNCF Sandbox project, and it is the cheapest way to find out whether this category helps your team at all.

• Scans for misconfigurations and failing workloads
• Explains findings with suggested remediation
• Runs as a CLI or in-cluster operator, and supports local model backends so cluster data need not reach an external API

Use case: run it against a misbehaving namespace and get a readable explanation of every failing object.

Where it does not fit: it analyses cluster state. It does not correlate with logs, traces, deployment history or cloud spend, and it fixes nothing.

Deployment: fully self-hosted. Pricing: free, Apache 2.0.

5. HolmesGPT and Robusta: best for alert-triggered investigation with a self-hosted alerting layer

HolmesGPT investigates when an alert fires, pulling the evidence a human would gather. Robusta is the alerting and automation platform around it.

• Alert-triggered investigation rather than on-demand scanning
• Pulls logs, events and related Kubernetes objects into one summary
• HolmesGPT is Apache 2.0 and self-hostable

Use case: a Prometheus alert fires at 3am and the investigation is attached to the notification before anyone opens it.

Where it does not fit: Robusta runs an in-cluster agent against its own SaaS control plane, and self-hosting the UI is an Enterprise option, so the open component and the platform have different deployment stories.

Deployment: HolmesGPT self-hosted; Robusta agent plus SaaS. Pricing: HolmesGPT free; Robusta not published.

6. Komodor: best for change tracking and guided Kubernetes troubleshooting across a team

Komodor's premise is that most Kubernetes incidents follow a change, so it builds a timeline of every change across the cluster and hangs troubleshooting off it. Klaudia is the AI layer on top.

• Change timeline across deployments, config and infrastructure
• Guided troubleshooting flows for common failure modes
• Klaudia AI investigates and explains

Use case: a service degrades and the timeline shows the config map that changed eleven minutes earlier.

Where it does not fit: the in-cluster agent feeds Komodor's SaaS backend where the AI runs, so cluster data crosses the boundary. Komodor sells autonomous remediation on both paid tiers; whether it is approval-gated by default is not stated publicly, so ask directly. Metoro is the nearest alternative for eBPF-based collection with no application instrumentation.

Deployment: in-cluster agent, SaaS backend. Pricing: 14-day trial, not published, per node.

AI for cloud cost, FinOps and Kubernetes capacity

Kubernetes cost and cloud-bill cost are different problems. The first is a resource-request problem inside your cluster, where the gap between what pods request and what they use is the waste. The second is a pricing-model problem outside it, solved with commitments and spot capacity. Our guides to Kubernetes cost optimization tools and AI FinOps agents cover each in depth.

7. Cast AI: best for automated Kubernetes rightsizing, bin-packing and spot handling

Cast AI continuously rightsizes workloads, packs them onto fewer nodes and moves eligible capacity to spot instances.

• Continuous pod rightsizing from observed usage
• Node bin-packing and instance selection
• Spot automation with fallback handling

Use case: a cluster provisioned for a peak that arrives twice a year is consolidated onto fewer, better-matched nodes without anyone editing a manifest.

Where it does not fit: the default connection is read-only, and write access requires an explicit permissions opt-in. Once granted, the agentic runbooks are approval-gated in Cast AI's own documentation; the broader automation is described without approval language, so confirm the default for your use case. ScaleOps is the closest comparable. Sedai takes a graded approach with three modes, defaulting to view-only, with unattended execution available only after opt-in and guardrail configuration.

Deployment: SaaS control plane with in-cluster agent. Pricing: trial available, not published.

8. nOps: best for AWS commitment and waste reduction with a FinOps agent

nOps attacks the pricing model rather than the resource requests, automating commitment purchases and spot usage on AWS.

• Commitment and savings plan automation
• Spot management for eligible workloads
• Clara agent for continuous optimisation

Use case: commitment coverage is managed continuously against actual usage rather than reviewed once a quarter.

Where it does not fit: it is AWS-native, so multi-cloud estates need something alongside it. Its automated rightsizing follows an opt-in-then-automatic pattern, with reversibility rather than per-action approval as the safety net. Kubecost, now owned by IBM, and OpenCost, which describes itself as a vendor-neutral specification for Kubernetes cost allocation, are the baseline. Vantage, CloudZero, Cloudchipr and Finout cover visibility, and Vantage also ships a FinOps agent with approval prompts.

Deployment: SaaS, AWS-native. Pricing: 14-day trial; not published, typically share of savings plus a spend-based fee.

AI in observability and AIOps

Two generations sit here. Correlation-first AIOps suites predate large language models and group alerts statistically. Newer tools reason over telemetry in language. Both are useful and neither replaces the other. Our guide to AIOps platforms covers the correlation generation, and AI alert investigation covers alert noise.

9. Datadog Bits AI: best for AI triage inside telemetry you already send to Datadog

If your telemetry is already in Datadog, Bits AI investigates without asking you to move anything or instrument anywhere new.

• Investigates alerts against existing Datadog telemetry
• Proposes remediation with a human approving the action
• No new agent to deploy if you are already a customer

Use case: an alert fires and Bits AI assembles related traces, logs and recent deploys into one narrative before anyone opens a dashboard.

Where it does not fit: the pricing unit is the trap. Bits AI is sold as AI Credits from $500 per month, on top of host and ingest licensing, and Datadog publishes log ingest at $0.10 per ingested or scanned gigabyte. The AI cost rides on a bill that already scales with volume, which surprises teams modelling it per seat.

Deployment: SaaS. Pricing: AI Credits from $500 per month, in addition to existing licensing.

10. Dynatrace Intelligence: best for deterministic causal analysis on a fully instrumented estate

Dynatrace, which now groups Davis AI and CoPilot under the Dynatrace Intelligence umbrella, derives causality from a dependency model of the estate rather than reasoning in language, which makes its answers reproducible.

• Deterministic causal analysis rather than probabilistic correlation
• Topology model built from its own instrumentation
• Natural-language querying layered on top

Use case: a latency spike is traced automatically to the specific service and change that caused it, with the dependency path shown.

Where it does not fit: the model is only as good as the instrumentation, so full coverage is the prerequisite rather than the outcome. Availability of the AI features on Dynatrace Managed is not stated publicly. New Relic AI is the nearest comparable; BigPanda and Moogsoft, the latter now inside Dell, represent the correlation-first generation.

Deployment: SaaS. Pricing: from $58 per month per 8 GiB host for Full-Stack Monitoring. No separate AI line item.

Running observability yourself: Prometheus, Grafana, OpenTelemetry and OpenObserve

You do not have to buy this layer. Prometheus and Grafana cover metrics and dashboards, OpenTelemetry is the default instrumentation standard, and OpenObserve offers a unified store for teams that want to own their telemetry and keep ingest costs predictable. Several tools above query these backends in place rather than re-ingesting, which matters when telemetry volume is what drives your bill.

AI for CI/CD, delivery and change risk

The interesting AI work in delivery is not writing the pipeline, it is judging the change before it lands. See how AI improves code reliability for the review side specifically.

11. Spacelift: best for AI-assisted Terraform plan review with policy as code

Spacelift orchestrates Terraform, OpenTofu, Pulumi and CloudFormation, and its AI layer splits in two. Saturnhead AI explains what a run did and why it failed. Intent, which Spacelift's own release dates to March 2026, takes a described outcome and proposes and applies the infrastructure change behind policy checks.

• Saturnhead summarises run failures in plain language
• Intent applies changes gated by policy as code
• Deployment options include self-hosted in your own cloud, on-premise and air-gapped, with private workers available even on SaaS

Use case: a plan fails on a policy violation and Saturnhead explains which rule tripped and what caused it, without anyone reading raw output.

Where it does not fit: the AI features send data to an external model in every configuration. The default routes to AWS Bedrock running Anthropic models, and the alternatives are Anthropic, OpenAI, Google Gemini, or your own Bedrock account. The Bedrock default keeps that traffic inside AWS rather than calling a vendor directly, which matters to AWS-committed teams, but it is still outside Spacelift's own boundary. If plan output cannot leave your infrastructure at all, confirm this before evaluating. Anyshift is worth knowing for blast-radius analysis before you ship a change.

Deployment: SaaS, self-hosted in your cloud, on-premise, air-gapped. Pricing: free tier of 2 users and 1 worker. Starter+ from $20,000 per year for unlimited users.

12. Harness AI DevOps Agent: best for teams that want AI folded into an existing delivery platform

Harness embeds AI across its platform rather than selling it separately, so the value depends on whether you already run Harness for CI, CD and feature flags. GitLab Duo, generally available as the Duo Agent Platform since January 2026, is the same motion on a different platform.

• Generates and troubleshoots pipeline configuration
• Error analysis on failed builds and deployments
• Worker Agents support bringing your own model; the DevOps Agent does not

Use case: a deployment fails and the agent reads the logs, names the likely cause and proposes the pipeline change.

Where it does not fit: the AI features are available through the Harness UI, and the Self-Managed Enterprise Edition module list does not include an AI module, so fully self-managed with AI is not something the documentation confirms. Pricing is not published.

Deployment: SaaS. Pricing: not published.

13. CodeRabbit: best for AI review on every pull request

CodeRabbit reviews pull requests line by line and posts findings as comments, which puts the AI at the point where a change is still cheap to fix.

• Line-by-line review with inline comments
• Learns from how a team responds to its suggestions
• Free for open-source repositories

Use case: a pull request touching Terraform and application code gets a first-pass review before a human opens it, so the reviewer starts from questions rather than from scratch.

Where it does not fit: it reviews diffs, so it sees the change and not the running system. It cannot tell you whether the change will behave under production load.

Deployment: SaaS. Pricing: free for open-source repositories; Lite from $12 per developer per month.

AI coding and IaC assistants

The most crowded bucket and the easiest to evaluate, because mistakes surface in review rather than in production. Cursor is also widely used here.

14. GitHub Copilot: best for in-editor code and pipeline config authoring

Copilot completes code inline and answers questions in chat. For DevOps work its most useful trick is generating Dockerfiles, Actions workflows and Terraform blocks from a comment.

• Inline completion, chat and pull request summaries
• Agent mode that can make multi-file changes and open a pull request
• Bring-your-own-key swaps the model backend, not the control plane

Use case: describe a build-and-push pipeline in a comment, get a working Actions workflow, then iterate in chat rather than in the docs.

Where it does not fit: it works on your repository, not your running estate. It has no view of production, so it cannot help during an incident.

Deployment: SaaS. Pricing: free tier of 2,000 completions and 50 chat messages per month; Pro $10 per user per month, Business $19.

15. Claude Code: best for repo-wide refactors of infrastructure code from the terminal

Claude Code runs in your terminal and operates across a whole repository rather than a single file, which suits infrastructure work where one change touches modules, variables and pipeline definitions at once.

• Agentic multi-file edits with a reviewable plan
• Runs against your choice of inference backend, including Bedrock and Vertex
• Configurable autonomy, from approve-every-edit to headless runs

Use case: migrate a Terraform module set to a new provider version, with the tool making the edits and you reviewing the diff.

Where it does not fit: it is a development tool and monitors nothing. Giving a headless agent write access to a repository deserves the same scrutiny as giving it cluster credentials.

Deployment: local CLI with your chosen inference backend. Pricing: $17 per user per month on Pro, billed annually.

16. Pulumi Neo: best for generating infrastructure code in a general-purpose language

Formerly Pulumi AI. Neo generates and refactors Pulumi programs, written in TypeScript, Python, Go or C# rather than a domain-specific language, so the model works in a language it has seen a great deal of.

• Generates infrastructure programs from natural language
• Works inside Pulumi Cloud alongside state and policy
• Free monthly token allowance

Use case: stand up a new environment by describing it, then keep the output under the same review and policy rules as hand-written code.

Where it does not fit: it is tied to Pulumi, so for a Terraform estate this is a migration decision rather than a tool decision. Ansible Lightspeed is the adjacent option on the configuration side.

Deployment: SaaS via Pulumi Cloud. Pricing: 5 million Neo tokens per month free, then $40 per month on Team plus credits.

AI for security and DevSecOps

17. Snyk: best for AI-suggested fixes to code, dependency and container vulnerabilities

Snyk finds vulnerabilities and proposes the fix as a diff, which is the part that matters. A scanner that produces a list creates work; one that produces a reviewable patch removes it.

• Snyk Agent Fix suggests patches for code issues
• Dependency, container and infrastructure-as-code scanning
• Snyk states it runs a custom, self-hosted model for fix generation rather than routing to a third-party LLM vendor

Use case: a transitive dependency picks up a CVE and Snyk opens a pull request with the version bump and reachability context.

Where it does not fit: it scans what you build, not what you run. Runtime posture is a different category, where Wiz, now part of Google Cloud, is the usual answer. Trivy is the self-hostable scanner baseline covering CVE and CIS benchmark checks, and is what NudgeBee orchestrates for those scans.

Deployment: SaaS. Pricing: free tier with capped monthly tests; Team from $25 per month per contributing developer.

Advisory or autonomous: which AI DevOps tools should get write access

Four questions are worth answering before an agent gets credentials that can change infrastructure, and most vendor documentation answers only the first.

1. What exactly can it change? Pods, node pools, cloud resources, source code, or all of it.
2. Is there an approval gate, and where does the approval arrive? A gate nobody sees during an incident is not a gate.
3. Is the action auditable and reversible? Can you show afterwards what changed and who agreed.
4. What is the blast radius? How many dependent services sit behind the thing it is about to touch.

Two caveats. Several tools move between rows depending on configuration, and the row that matters is the one your team will actually run in. And "unattended after opt-in" is not reckless: in each case it requires a deliberate decision and guardrails.

Can you self-host it? Deployment model, data egress and bring-your-own-model

SaaS-only means the vendor runs everything and your data goes to them. BYOC means the vendor's software runs in your cloud account but the control plane is usually still theirs. Fully self-hosted means you run all of it and nothing leaves unless you wire it up.

Of the 17 tools here, three run fully inside your boundary: K8sGPT, HolmesGPT and NudgeBee. Spacelift offers self-hosted, on-premise and air-gapped deployment for orchestration, though its AI features call an external model in every configuration. Everything else is SaaS or an agent reporting to a SaaS backend.

Bring-your-own-model is a separate axis. Pointing a SaaS tool at your own LLM endpoint stops prompts reaching a model vendor, but the tool itself still sees your data. Harness supports BYOM on Worker Agents but not on the DevOps Agent. Spacelift lets you choose the provider, all external. NudgeBee supports 11 provider routes including local options like Ollama and vLLM. Copilot's bring-your-own-key changes the model, not the control plane.

If your constraint is regulatory rather than preferential, that distinction is the one to interrogate. Our guide to open-source incident investigation tools covers the genuinely open options.

What these tools cost, and the pricing unit that surprises teams

Nine of the seventeen do not publish a price at all. Of those that do, the unit varies enough to make comparison hard: per seat for coding assistants, per contributing developer for Snyk, per developer for CodeRabbit, per node for Komodor, per host for Dynatrace, per cluster for NudgeBee, and per gigabyte ingested underneath Datadog.

The ingest unit is where budgets break. AI observability features are usually priced on top of a bill that already scales with telemetry volume, so adopting AI investigation on a high-volume estate increases spend on two axes at once.

Share-of-savings pricing, common in FinOps, aligns incentives neatly but makes cost hard to forecast, because a successful year is an expensive one.

What AI still cannot do for your DevOps team

It does not know what your business considers important. A tool can tell you a service is degraded; it cannot tell you that this one processes payments and that one renders avatars.

It does not know your ownership model. The fastest part of most incidents is knowing who to ask, and that lives in people's heads and in a wiki nobody updates.

It can be confidently wrong. A remediation suggested with complete assurance and no supporting evidence is more dangerous than no suggestion, which is why visible reasoning matters more than a fast answer.

And a symptom is not a cause. A 503, a CrashLoopBackOff, a failed health check: a tool can name any of these in seconds and none of them is a root cause.

Common mistakes when adopting AI in DevOps

Adopting agentic features before the telemetry is clean. An agent reasoning over noisy or mislabelled data produces confident nonsense faster than a human would.

Treating an agent as autopilot. The tools that work in production keep a human in the decision, particularly early while you are calibrating trust.

Buying a point tool for something your platform already ships. Several vendors here include AI features in tiers you may already pay for.

Skipping review on generated infrastructure code. Generated Terraform looks correct more reliably than it is correct.

Buying roadmap instead of shipped capability. Ask what is generally available today. PagerDuty's fully autonomous responder is early access, not GA, and that matters if it is why you are buying.

See faster DevOps incident recovery for the operational side.

How to choose: a six-question decision path

1. What breaks most often? Code quality, delivery, or production behaviour. Buy for that, not for the category.
2. Can your data leave your environment? If not, the shortlist is short, and you should establish that first rather than last.
3. Do you want advice or action? And if action, who approves it and where.
4. What do you already pay for? The cheapest AI capability is often the one bundled into a platform you run today.
5. What is the pricing unit? Model it against your telemetry volume, node count or seat count before the trial ends.
6. Is it shipped? Ask what is generally available rather than announced.

Frequently asked questions

Where this is heading

The interesting shift in this category is not model quality, it is the governance question. As more of these tools move from explaining to acting, the durable difference between them becomes what they are allowed to change and who approves it. Read our view on the future of DevOps.

If your constraint is that production telemetry cannot leave your environment, NudgeBee is free to self-host on up to two clusters.

Methodology, sources and update log

Every tool was verified against the vendor's own site or documentation on 19 August 2026. Pricing was checked the same day and is stated with its unit. Where a vendor does not publish a price, deployment detail or approval behaviour, the entry says so rather than estimating.

Changelog

• 19 August 2026: first published. Recorded during research: Pulumi AI is now Pulumi Neo; Dynatrace groups Davis AI and CoPilot under Dynatrace Intelligence; Amazon Q Developer is being retired and replaced by Kiro, a separate product rather than a rename; Zenduty is now Xurrent IMR following acquisition; Wiz is part of Google Cloud; Kubecost is owned by IBM; Moogsoft is inside Dell; GitLab Duo reached general availability as the Duo Agent Platform in January 2026; Spacelift Intent reached general availability in March 2026; Snyk's fix capability is now Snyk Agent Fix; PagerDuty's Fully Autonomous Responder remains early access.

Reviewed quarterly. Next review November 2026.