Infrastructure agents · on duty 24/7
An infra agent that never clocks off. An army of them, if you like.
sudoX is a crew of agents for your infrastructure, awake around the clock. One watches your logs and metrics and learns what normal is. One investigates when it stops being normal. One drafts the fix. One runs it — after a person says yes — and one comes back later to prove it worked. Add accounts and they scale out; you keep the gate.
Try the loop
The whole product, in one strip.
This is what the worker does on every tick, compressed. It reads log and metric sources, learns what normal looks like, notices when it stops — and only in detect does it open an incident, investigate with read-only tools and draft a plan. The plan stops at a gate. You are the gate. Turn the dial and watch what changes; when the blue card appears, decide.
what the tape is
The worker's own log
Each line is a thing the product records: a signal read, a baseline learned, a classification, an audit entry. In the console the same lines are the Logs, Analyses and Audit pages, kept.
what the dial is
The lifecycle mode
training watches and learns. shadow judges and records what it would have raised, so you can grade it before trusting it. detect is the only mode that reaches people. It is set per deployment and narrowed per environment.
what the card is
The approval gate
A plan never runs without a named person releasing it. Approve, and the audit line at the end carries your name; reject, and nothing ran and the finding stays open. There is no third outcome.
Four things the product is built around
The rules that make its numbers mean something.
Nothing changes without a person.
Reading is autonomous. Everything that writes stops at a gate. An AI never touches infrastructure directly — it produces a plan, an inspectable artifact of ordered steps, each carrying a risk class, a target and how to undo it. What the planner wrote is exactly what a named operator approves and exactly what runs. There is no gap between the three for something to slip through. Build a plan and watch the gate decide.
A claim is not a result.
A plan completing means work happened, not that it worked. Only when the estate is read again and compared against what was recorded when the problem was found does a finding become verified — and only verified outcomes count as impact. A dry run never counts as an execution, a rollback never counts as a success, and a problem that comes back is recorded as a regression rather than quietly reopened. Walk a finding through it.
A more specific setting narrows, never widens.
An environment's agent mode can only narrow the deployment's. A credential's role inside an environment can only narrow its organisation role. That one rule is what keeps the org-wide switch a working emergency brake: dialling production down to shadow stops it being paged about and stops it being repaired unattended, because both read the same lever. Turn both dials.
It says what it does not know.
An estate never discovered, a bill never read, an environment nobody assessed and an environment with nothing wrong look identical if a page only renders what it has. So every surface names the gap: "not assessed yet" instead of a reassuring zero, "part of this could not be read" instead of a smaller number, "on, but not running" beside a source whose credential expired. Take a permission away and read what it says.
The loop, page by page
Where each step lives in the console.
The console is organised by the job somebody is doing, not by the module that does it. The three-in-the-morning zone comes first.
| Step | What happens | Console page |
|---|---|---|
| watch | Log and metric sources are read every tick; a catalog of patterns and a baseline per signal are learned. | Logs · Watching |
| notice | A pattern never seen, or a signal outside its band, is classified. In shadow it is recorded; in detect it becomes an incident. | Now · Incidents · Decisions |
| investigate | An analysis agent uses read-only tools — describe, query, history — and writes a hypothesis with the calls it made. | Analyses · Ask |
| plan | A remediation planner drafts ordered steps with a risk class, a target and a rollback each. The plan is an artifact. | Actions |
| gate | The approval policy decides who must say yes: nobody, one named person, or two. Two floors no rule can lower. | Actions · Approval policy |
| execute | Step by step, through the same tool surface, with the policy re-evaluated at the moment of execution. | Actions · Audit |
| verify | The scope is read again and compared with the evidence recorded at discovery. A result, not a claim. | Findings · Impact |
| remember | An episode is recorded, a runbook drafted, knowledge proposed — never written until a person accepts it. | Runbooks · Knowledge graph |
| sweep | On a schedule, the estate is inspected scope by scope for problems nobody has noticed yet. | Infrastructure · Findings · Drift · Cost |
What you get
From a connected cloud to a watched one, without typing log group names.
Connect
A wizard that tests before it stores
Cloud accounts and databases, per organisation, every credential sealed server-side. A credential is probed before it is saved, so you find out here rather than during an incident — and the first look starts the moment it is stored.
See the first look →Discover
The estate, drawn
A compiled-in, read-only probe list draws each account as a typed graph. A probe that fails is shown as a gap, because an unreadable subsystem must not look like an absent one. Every snapshot is kept, so drift is a diff, not an inference.
Walk the graph →Derive
What to watch, proposed
From what discovery found, a reviewed catalog proposes the log sources and golden-signal metrics — one query per kind of resource, so it keeps covering the estate as it changes. You review a list and turn things off.
See the proposal →Assess
Findings with evidence
On a schedule, the estate is inspected scope by scope. Each finding carries the observation, the evidence, the recommended fix and how to prove the fix worked. Re-running updates what exists instead of duplicating it.
Walk a finding →Act
Plans, gated by policy
Reads run alone. Writes wait for a named approver. Destructive steps and writes with no rollback always need a human, whatever any rule says. Where an environment is armed, cleared plans run unattended — three a pass, oldest first.
Run a pass →Prove
Verified, on a schedule
"Resolved" turns into "verified" without anybody remembering to check. The impact report counts only measured before-and-after pairs, never an estimate.
What happens after the fix →Fourteen things you can operate
Every mechanism, as a thing to try.
Each of these runs the same rule the product runs, in the same words. None of them is a video.
The loop
Turn the mode dial; be the gate.
The baseline
Watch a metric learn what normal is, then break it.
The pattern catalog
Send a log line it has never seen.
Claim and result
Walk a finding from open to verified — or to regression.
Alert routing
Switch rules and see where each page lands, or is held.
Narrow, never widen
Two dials, one effective mode, and the brake that stays a brake.
The gate
Build a plan; watch the two floors hold under every policy.
Derivation
Choose what discovery found; read what is proposed and what is not, and why.
The first look
Connect a cloud; watch four phases run once, and refuse to run twice.
The autonomous runner
Arm an environment; run a pass; read the audit lines.
Who can do what
Pick a role, an assignment and an action; read the verdict and its audit line.
The reorder window
Late and duplicate log entries, caught or counted.
The knowledge graph
Click a database; see what depends on it, and in which estate.
What must be granted
Take a permission away and read exactly what the console says.
Plugs into
Mostly through MCP, so a new system is a config entry rather than a code change. Models are equally swappable — OpenAI, Anthropic, Gemini, DeepSeek, Qwen, or something self-hosted through Ollama — and a different model per task, changed without a restart.