checkout-api runs on ECS · prod · eu-west-1 and depends on orders-db (Aurora PostgreSQL) and payments-gw. Checking deploys, flags, database and dependencies…Custom autonomous systems · Mock demo
When an alert fires, First Mate works out what's affected, gathers evidence from your monitoring, deploys and feature flags, ranks the likely causes, and posts a first evaluation to the team's channel. After recovery it drafts the post-mortem. It reads production; it never changes it on its own.
The metric at the centre: time to restore service (MTTR). Everything it does is judged by how fast the error rate it is watching comes back down.
How far to trust an AI with an incident, and where it falls short, is covered in our guide AI root cause analysis for incidents: what it can and can't do.
Mock demo · fictional company and data
Press Restart, then Next (or →). The run stops when it needs your approval: you play the human.
checkout-api runs on ECS · prod · eu-west-1 and depends on orders-db (Aurora PostgreSQL) and payments-gw. Checking deploys, flags, database and dependencies…Order placement on the web shop and the mobile app. Browsing, search and the basket are healthy. Payments are not failing; requests time out before they reach the payment step.
recs-at-checkout flag, rolled to 100% at 09:41, added a slow query to every checkoutHigh▸recs-at-checkout off. Reversible in one click, affects only the recommendation strip.order_items(sku_id) before re-enabling. Draft migration attached as PR #2291.recs-at-checkout off (approved by you, on call). Watching the error rate, database CPU and connection pool. I'll confirm when checkout is back to normal; if it isn't within 10 minutes, I'll say so and look again.For 17 minutes, about 14% of checkout requests timed out. Rolling the recs-at-checkout flag to 100% added an unindexed query to every checkout, saturating the orders database.
The recommendation query filters order_items by sku_id, which has no index. At 10% of traffic the load was absorbed; at 100% it was not.
Drafted by First Mate from the evidence above. A human edits and merges it; nothing is published automatically.
Figures describe this mock scenario. They are design targets, not measured results.
For many teams, an off-the-shelf AI SRE tool is enough, and it's the right place to start. These are the ones we'd look at first, described from their own sites (checked October 2026).
Investigates alerts as they fire, forms hypotheses from Datadog telemetry and suggests code fixes. Datadog's docs now call it Bits Investigation. Enough when most of your metrics, logs and traces already live in Datadog.
An AI SRE that starts investigating when an incident is declared, plus AI post-mortem drafts. The only change it makes to your systems is a pull request you merge. Enough when you want incident management and AI investigation in one tool.
AI SRE that ranks evidence-backed hypotheses with confidence scores and suggests fixes, with human sign-off on every change. Enough when you want this design, ready-made, alongside incident management.
An SRE Agent for triage, diagnosis and approved remediation, and a Scribe Agent that captures incident conversations for post-incident reviews. Enough when PagerDuty already runs your on-call.
AI incident summaries and AI-assisted retrospectives. Freshworks has announced it is acquiring FireHydrant. Enough when your main gap is write-ups and stakeholder updates rather than investigation.
Positions itself as an AI SRE that goes on call, triages alerts and mitigates incidents. Enough when you want a more autonomous agent and are comfortable with a vendor running it.
More on the trade-offs in our guide to AI root cause analysis for incidents, and on how we build these in Custom Autonomous Systems Engineering.
Often for incidents that follow a recent change or match a known pattern, such as a feature flag rollout that adds a slow query. Much less reliably for novel or multi-cause failures, where models tend to anchor on the first plausible story. That's why First Mate ranks several causes with the evidence behind each and lists what it ruled out, so the on-call engineer can check its reasoning in a minute.
Not by default. First Mate has read-only access and every fix it suggests waits for a person to approve it. Once a team trusts it with one specific action that is verifiable, reversible and contained, that single action can be allowed to run on its own.
It watches the metric that triggered the alert, plus the signals behind the suspected cause, until they are back to normal. If the error rate hasn't recovered within an agreed window after the approved fix, it says so and reopens the other hypotheses instead of closing the incident.
It can write a solid first draft. First Mate builds the timeline from timestamped evidence such as metrics, deploys and flag changes rather than from chat, keeps it blameless, and marks action items as suggestions. The draft is opened as a pull request so a person edits and merges it.
Read-only access to your alerting and monitoring (for example Datadog, CloudWatch or Prometheus), logs and traces, deploy history, feature flags, your service catalog, and the chat tool where your team handles incidents. Where it couldn't check a source, it says so in the evaluation.
No. It's triggered by the alerts and paging you already have, and posts into the channel your team already uses. It adds the investigation, the recovery check and the post-mortem draft on top of your existing on-call process.
Time to restore per incident is the headline, but averages are misleading because incident durations are skewed. Also track what the agent directly controls: time from alert to first evaluation, and how often its top-ranked cause turned out to be right. Our guide covers measuring MTTR honestly.
First Mate has read-only access to production. Every change it suggests, even a reversible one like turning off a flag, waits for a person to approve it, and after acting it watches the same metric to prove the fix worked. When you trust it with a specific action, you can let that one action run on its own: the operations that are verifiable, reversible and contained, one at a time.
We build it around your monitoring, your services and your on-call process. Tell us what wakes your team up at night.
More demos