Solutions · Turn alerts into fixes
AI incident remediation: from a production alert to a reviewed fix.
A hosted alert or a new error reaches Autopilot on-call with no URL or token to wire. It files one deduplicated issue, investigates read-only and writes up the likely cause and a fix. A person decides whether to work it. Then an agent writes the fix and opens a pull request for normal review.
The problem
Why alerts rarely reach the code that caused them.
An alert fires in one tool, the error lands in another, the deploy history lives in CI and the code is in GitHub. Whoever is on call spends the first half hour joining them up, and an AI assistant bolted onto any one of them sees a single slice.
Wiring an agent into that chain means webhooks, tokens, routing rules and de-duplication you write and maintain, and an agent that files a hundred copies of one issue the first time an alert flaps.
How it works
How it works on Reliant.
Two linked steps with a person between them: on-call investigates and files, a person approves, the issue worker fixes. It is not one auto-fix workflow, on purpose.
- Forge · deterministic
An alert fires, or an error is new
Forge apps ship with default alerts on error rate, latency, crash loops and failed deploys. Errors are grouped into issues, which raise an alert when one is new, comes back in a later release, or spikes to five times its previous day.
- Reliant · deterministic
It reaches the agent with no wiring
A workflow trigger of webhook: {source: alerting} receives your organization’s critical and error alerts on Reliant hosting. No URL, no token, no route to write, and your other receivers are unchanged. A silence mutes it like any receiver.
- Agent
On-call files one issue and investigates
Autopilot on-call files one GitHub issue per new alert or error, in the repository that owns the service; a repeat adds a comment, never a second issue. Working read-only from the alert, the grouped error and your code, it writes up severity, evidence, the likely cause and a proposed fix.
- You
A person approves the work
On-call labels the issue agent:triage and stops. If you agree with it, add agent:ready. Nothing is fixed until someone does.
- Agent
The issue worker writes the fix
It writes a failing test and the fix in its own workspace, runs your fast checks in a shell with no credentials, pushes an autopilot/* branch and opens a pull request that closes the issue.
- You
Normal review, then a new release
PR review and security review run, the responder answers comments, and a person merges. The fix ships as a new release, and its marker appears on the dashboard where the alert fired.
# your org’s alerts; no URL or tokentriggers: - name: alertmanager webhook: {source: alerting}Who does each step
- Forge · deterministic
- Reliant · deterministic
- Agent
- You
The evidence
What an agent can read, and how.
The built-in investigator works from the alert, the grouped error and your code. Agents you point at an incident yourself reach the rest through Forge.
The alert and the error
The alert’s labels, or the error’s service, environment, file and function, and the build it was first seen in, with a link to it in the console.
Logs, traces and profiles
forge obs gives an agent logs, traces, metrics, errors, sessions and profiles as plain commands with JSON output. The built-in investigator gets them only when you give it a tool for them; otherwise its write-up says what it could not see.
Read-only production data
forge db query runs in a read-only, audited transaction with a timeout, and the read-only role cannot see credential columns. It needs the db:query grant.
The release that caused it
Errors are tied to the release they first appeared in, and forge obs check names the release that regressed.
What you get
What you get.
Alert-triggered auto-fix work, with the approval in the middle where it belongs.
Alerts to an agent, zero wiring
One trigger line connects a workflow to your organization’s alerts and new, regressed or spiking errors.
WebhooksAutopilot on-call
De-duplicated issues routed to the right repository by service name, each investigated read-only.
Fix featuresDefault alerts and receivers
Error rate, latency, crash loops and failed deploys, routed to Slack, PagerDuty, email or a webhook. Alert rules are code you can run locally.
Dashboards and alertsErrors grouped into issues
Fingerprinted, symbolicated and tied to the release that introduced them.
Run featuresTelemetry agents can query
Logs, traces, metrics, errors, sessions and profiles from the command line, with JSON for agents.
Observability for agentsThe fix loop
Issue worker, PR review, security review and review responder take the approved issue to a pull request a person merges.
Autopilot guideRelease verdicts
After the fix deploys, forge obs check judges the new release against the window before it.
Release featuresAny trigger for your own runbooks
Schedules, webhooks, GitHub, Slack, Gmail and Twilio events, or another run finishing, can start a workflow you write.
Automations
Why one platform
Why this takes one platform.
The alert, the error, the release and the code usually live with four vendors, and joining them is the slow part of every incident. Here the platform that runs the app also groups its errors, records its releases and runs the agents, so an alert arrives already tied to a service, a repository and the version that introduced it, and the hand-off to an agent needs no webhook, token or routing rule from you. A bolt-on tool can only see the slice it was wired into.
Honest limits
What it does not do, and what stays yours.
Remediation stays reviewable at every step.
It is two linked steps, not one auto-fix: on-call files and investigates, a person approves, the issue worker fixes.
There is no automatic rollback. Reliant rolls forward to a new release, or you redeploy an earlier one; a rollback floor stops anyone deploying a release older than a breaking migration.
The built-in investigator reads the alert, the grouped error and your code. Logs and traces reach it only through a tool you give it.
By default at most five new alerts are investigated per notification; the rest are filed without an investigation.
Zero wiring is for environments on Reliant hosting. Self-hosted, on-call is a plain webhook with a URL and token you add to Alertmanager.
Alerts map to repositories by service name. An alert nothing maps goes to a default repository if you set one; otherwise it is reported in the run and nothing is filed.
Questions
What people ask first.
Does the agent fix production automatically?
No. Autopilot on-call files and investigates an issue, labels it for triage and stops. A person who agrees adds the ready label; only then does the issue worker write the fix, and the pull request goes through normal review before anyone merges it.
Will a flapping alert open a pile of duplicate issues?
No. A repeat of the same firing alert within four hours is one event, and every alert or error carries a key that is searched for among open issues, so a match gets a comment instead of a second issue. If the search fails, nothing is filed: a missed alert is cheaper than a duplicate.
Does it roll back a bad release?
No. Reliant rolls forward: the fix ships as a new release, or you redeploy an earlier one yourself. A failed deploy never replaces what is serving, and a rollback floor stops anyone deploying a release older than a breaking migration.
What does an investigation cost?
One investigation of up to 45 agent turns per new alert or error, at most five per notification by default, and nothing for a repeat. All three numbers are settings you can lower.
Keep reading
Related.
- Start a new app
A production-shaped app from the first commit, deployed by your agent.
- Bring an existing app
Any language, any repo: environments, releases and monitoring without a rewrite.
- Automate PR review to merge
Agents review every pull request and answer comments; gates decide what merges.
- Agents & Workflows
Multi-agent workflows in YAML, Autopilot and triggers, on any model.
- Observability
Traces, logs, metrics, alerts and release verdicts, wired from day one.
- Why Reliant
One platform that knows every seam, so your agents can too.
- All features
Every capability by stage of the lifecycle, each with its status.
Give on-call an agent.
Turn on Autopilot on-call in Reliant. Nothing is fixed without a person’s yes.