Agents own the analysis.
Humans own the outcome.
Bits AI is Datadog's family of agents. One investigates alerts, one answers questions, one writes code fixes, one triages security signals, one builds other agents, and the newer ones from DASH 2026 detect, remediate and optimise. Each is useful. None of them should decide alone. We scope what each agent may read and touch, set the guardrails, and review every month what it proposed and what an engineer did with it.
Quick facts
Related
- Bits Investigation, the agent on the alert
- Bits Agent Builder, agents for your own runbooks
- Datadog for AI, the other half: observing your own AI
One family, several jobs, and a different human decision behind each
The alert fires at 3am. By the time the engineer has found their laptop, an agent has read the trace, the deploys and the logs and has a hypothesis waiting. That is useful. It is only useful if the engineer knows what the agent read, what it was allowed to do, and that it did nothing yet. The agents are different and so are the decisions they leave to a person.
Investigation and detection read, then propose. Bits Investigation takes an alert and returns a root-cause hypothesis with the evidence. Bits Detection, in Preview, reads production traffic and creates managed monitors for critical services. Both act on telemetry. Neither changes your runtime.
Remediation and Code propose a change. Bits Remediation, with capabilities in Preview, turns an investigation into a next step: a message, a ticket, or an infrastructure action behind a guardrail. Bits Code turns a finding into a pull request. The change exists. A person approves or merges it.
Chat and Agent Builder work inside your permissions. Bits Chat answers questions with the access of the person asking, and no more. Bits Agent Builder lets you compose agents from the Action Catalog, and an agent assigned to a work item runs with the permissions of whoever assigned it.
Security Analyst recommends a verdict. Bits Security Analyst, generally available, investigates Cloud SIEM signals and labels each one with a conclusion and the evidence. The response action is taken by a person from the side panel. We cover it on its own page, without the jokes.
Every agent on, no guardrail set, a credit budget nobody owns, and a hypothesis mistaken for a finding
How it bills. Bits Chat, Bits Investigation, Bits Code and Bits Agent Builder draw on AI Credits, and Datadog prices Bits Database Optimization in AI Credits too. Datadog sells credits on demand or in monthly bundles. Bundle credits reset on the first of each calendar month and do not roll over, and usage above the bundle is billed at the on-demand rate. Each product consumes credits per unit of work: per chat message, per investigation, per code fix, per agent run. Datadog publishes an estimate for each, for reference only. Bits Security Analyst is metered differently: Cloud SIEM includes a monthly allotment of its investigations that scales with the volume analysed. Datadog's pricing does not list Bits Detection or Bits Remediation separately, so how those bill is confirmed with Datadog at the start of each engagement. Administrators can cap credit usage for the whole organisation and per user. We publish no rates. The meter is credits, and credits are consumed by whatever the agents are allowed to do, which is why scope is the first decision.
The configuration traps. Every agent is enabled from the trial because the toggle was there. Bits Remediation guardrails are left at the default, so the question of whether an agent may restart a pod is answered by whoever set up the connection. Bits Code is given the whole GitHub organisation rather than the services that matter. The knowledge sources are empty, so the agents reason without your runbooks and propose what a generic estate would need. Credits run out in the third week of the month and the agent that was useful stops. The hypothesis from an investigation is pasted into the postmortem as the cause, without the step where somebody checked. None of this is the agent failing. It is an agent doing exactly what it was permitted to, with nobody having decided what that should be.
Field noteThe most common gap we see is the agent nobody scoped. It was enabled to try it, it worked well enough to keep, and six months later it has access to every repository and every environment because the trial scope was "everything".
Field noteThe pattern we keep finding is the investigation summary that becomes the postmortem. The hypothesis was plausible and well written. The verification step, where an engineer confirms the cause against the data, was the step that got skipped because the summary looked finished.
Field noteCredits are a shared pool with no owner. An organisation-wide cap goes on to protect the budget, one team's agent consumes most of it by the second week, and another team's on-call engineer discovers at 3am that the investigation agent has gone quiet.
Field noteThe knowledge sources were never filled in. The agents reason from telemetry alone and propose the fix a generic estate would need. The team stops reading the proposals, and the agent that could have helped is ignored on the night it matters.
Decide what each agent may read, what it may propose, and who says yes. Then switch it on.
We start with a Datadog HealthScan, read-only and fully automated. It reports which Bits agents are enabled, what connections and repositories they can reach, how the guardrails are set and what the credit consumption has been. The scoping then runs as part of Managed Observability onboarding, one agent at a time.
- Inventory the agents and their reach. Every enabled agent, every connection it can use, every repository Bits Code can see, every rule Bits Security Analyst watches. Written down before anything changes.
- Set the guardrails to Ask or Deny. For each integration action Bits Remediation can take, by environment, service and resource tag: Ask a named approver, or Deny so the agent only recommends. Production starts at Deny and earns Ask.
- Scope the agents that touch code and security. Bits Code limited to the repositories, services and teams that matter. Bits Security Analyst scoped by rule and minimum severity. Work-item agents assigned only by people whose permissions you would give the agent.
- Fill the knowledge sources. Runbooks, architecture notes and organisation context go into the Skills Library and knowledge sources. The agents then propose what your estate needs, not what a generic one would.
- Set a credit budget with an owner. A monthly credit allotment per agent or per team, a monitor on consumption, and a named person who decides what happens when it is reached.
- Hand over the agent register. Each agent with its scope, its guardrails, its knowledge sources, its budget and the human who owns its outcomes. The register is what the board asks for, and what we review monthly.
Access is admin on the Datadog organisation for the agent, guardrail and credit settings. Scoping the agents' own connections to your cloud accounts and repositories is done with you, using your accounts, so no credential of yours is held by us.
An agent that was safe in June has new capabilities by September
The agents change faster than any other Datadog product: features leave Preview, new actions appear in the catalogue, a guardrail default moves. Managed Observability reviews every agent every month against what it proposed and what a human decided. Where an agent proposes an action in the runtime we operate, Critical Support is the human who decides.
Every month, proposals meet decisions. What each agent proposed, what an engineer accepted, changed or rejected, and why. An agent whose proposals are mostly rejected is rescoped or its knowledge sources are fixed. One that is mostly accepted earns a wider Ask.
Scope follows the estate and the release notes. New repositories, services and environments are checked against each agent's reach. New agent capabilities from Datadog are reviewed before they are allowed, not after they have acted.
The guardrails are kept honest. Every guardrail and every approval is re-read quarterly against the incidents that happened. A Deny that blocked a fix an engineer would have made is discussed. An Ask that nobody answered is fixed.
The outcome is ours when the runtime is. Under Critical Support our on-call engineer is the approver for any action an agent proposes in the runtime we operate, 24x7. The agent writes the hypothesis. The engineer verifies it, acts, and signs the record.
We operate the stack. You own the product.
We run Datadog's agents inside our own operations, every day, under the same rules we set for customers: Agents own the analysis. Humans own the outcome. That is not a slogan about AI caution. It is a description of who signs the postmortem. The agents, their scopes and their records stay in your Datadog, and you can take the operation in-house whenever you choose.
Datadog is our proof, not our headline. Datadog certified one managed provider globally. We are it. The Powered by Datadog accreditation means Datadog's own engineers reviewed how we instrument, alert, respond and report. We are also a Datadog Advanced Partner, ISO 27001 certified and Cyber Essentials Plus certified, with offices in Cardiff, London and Dublin.
Datadog Bits AI, answered plainly
Which Bits AI agents exist, and which are in Preview?
As of DASH 2026 the family is Bits Investigation, Bits Chat, Bits Agent Builder, Bits Code, Bits Security Analyst, Bits Detection, Bits Remediation and Bits Database Optimization, plus Bits Data Analysis for warehouse queries. Datadog states that Bits Chat, Bits Agent Builder, Bits Code and Bits Security Analyst are generally available and that Bits Detection is in Preview. Bits Remediation includes capabilities in Preview, among them one-click infrastructure actions and the guardrails that govern them. Bits Database Optimization is available for PostgreSQL and MySQL and in Preview for SQL Server. The status of each agent changes often, and we confirm it against Datadog's documentation at the start of every engagement rather than quoting it from memory.
How is Datadog Bits AI billed?
Bits Chat, Bits Investigation, Bits Code, Bits Agent Builder and Bits Database Optimization bill through AI Credits. Datadog sells credits on demand or in monthly bundles that reset on the first day of each calendar month and do not roll over, and each product draws on the pool per unit of work: a chat message, an investigation, a code fix, an agent run. Bits Security Analyst investigations come with Cloud SIEM as a monthly allotment. Administrators can cap credit usage for the organisation and per user. We do not publish prices. We set a credit budget per agent or per team with a named owner, monitor consumption, and treat an agent that goes quiet mid-month as a configuration problem to fix.
Can a Bits AI agent change our production environment?
Only if a guardrail allows it. Bits Remediation can recommend an infrastructure action such as scaling a deployment or restarting a pod, and in Preview can run it with one click. The guardrails decide, per integration action and per environment, service and resource tag, whether the agent must Ask a named person first or is Denied and can only recommend. Bits Code opens pull requests and engineers decide what to merge. Bits Chat and work-item agents act within the permissions of the person using or assigning them. We set production to Deny by default and widen it deliberately.
What does "Agents own the analysis. Humans own the outcome." mean in practice?
The agent reads the telemetry, forms the hypothesis and drafts the proposal, and it does that faster and more consistently than a person at 3am. A named engineer verifies the hypothesis against the data, decides whether to act, takes or approves the action, and signs the record. The agent register lists, for every agent, which human owns its outcomes. Under Critical Support that human is our on-call engineer for everything in the runtime we operate.
What does Critical Cloud do with Bits AI every month?
Under Managed Observability we review each agent: what it proposed, what an engineer accepted, changed or rejected, and why. Agents whose proposals are mostly rejected are rescoped or given better knowledge sources. New repositories, services and environments are checked against each agent's reach, and new capabilities Datadog releases are reviewed before they are permitted. Guardrails and credit budgets are re-read against the incidents that happened. Under Critical Support our engineer is the approver for any action an agent proposes in the runtime we operate.
- Datadog products: every product, and how we implement and run each one
- Bits Investigation: the root cause proposed, verified by an engineer
- Bits Security Analyst: signals triaged, with a human on the decision
- Agent Observability: observing the agents you build, including these
Agents that propose. Engineers who decide.
Tell us which Bits agents are switched on today. We will show you what each one can reach, what it has proposed, and what the guardrails should say before the next one leaves Preview.