Custom agents

Event-driven and scheduled agents, each defined by a trigger, a plain-English instruction, and a destination.

Most teams eventually script the same checks: an hourly sweep for unused resources, a daily digest of slow queries, a review pass nobody has time to run. Sherlocks executes those jobs, keeps them working, and reports into the tools your team already uses.

When?Every deploy, just after rollout

Watch the canary

Flag any anomaly in the golden metrics while the canary is still at 1%.

Lands in #deploys in Slack

Where the cost actually sits

The individual checks are rarely difficult to write. An hourly sweep for unused resources, a digest of the slowest queries, a review pass over incoming changes: most teams have written some version of each.

Authoring the script is seldom the constraint. What determines whether it is still useful a quarter later is who keeps it running, whether its output stays accurate enough to act on, and whether it reaches anyone.

  • Execution

    The trigger fires without a scheduler to maintain, a host to keep available, or credentials to rotate on a schedule.

  • Reliability

    When a check fails for its own reasons rather than yours, resolving it is our responsibility. Automations that lose credibility get muted, and a muted check provides no coverage.

  • Reporting

    Findings are delivered where the work already happens: on the pull request, in the relevant Slack channel, or on the incident record.

Agents teams are running

Four examples, each shown with the instruction the agent was given.

  • Every hour

    Idle resources

    Sweep the cluster for resources nobody is using: pods sitting far under their requests, PVCs with no pod attached, load balancers with no traffic. Post the ten most expensive, with what they cost per week.

    Capped at ten and ranked by cost, so the output stays short enough to act on.

    Lands in #platform in Slack

  • A pull request is opened

    Index coverage for new queries

    When this pull request adds or changes a query, check whether the tables it touches already have an index that covers it. If they do not, say so on the pull request, before it reaches production.

    Lands in Comment on the pull request

  • Every 24 hours

    Slow query digest

    Every morning, list the slowest statements of the last 24 hours ranked by total time, and tell me what changed about each one since yesterday.

    Ranked by total time rather than worst single run, so a frequent query outranks a rare outlier.

    Lands in #sre-daily in Slack

  • A pull request is opened

    Build compatibility on dependency bumps

    When a PR gets opened in data-service, check whether the changes can break the build. The go.mod gets bumped to a new Go version and the Dockerfiles don't, so the build fails.

    Close to verbatim from a customer request.

    Lands in Comment on the pull request

How an agent is created

You provide the description. Sherlocks requests only the details it cannot infer from it, then returns a specification for review before the agent runs.

  1. Description

    “When a PR gets opened in data-service, check whether the changes can break the build. The go.mod gets bumped to a new Go version and the Dockerfiles don’t, so the build fails.”

  2. Clarification

    When the agent should run, and where its findings should go. Limited to what the description leaves open.

  3. Approval

    Version one of the specification is returned in full: trigger, job, and destination. Nothing runs until it is approved.

  4. Deployment

    Trigger set to a GitHub pull request being opened, findings delivered as a top-level comment on that pull request.

Predefined agents are available ready to configure, or you can define your own.

Popular custom agents

  • Every deploy, just after rollout

    Watch the canary

    Flag any anomaly in the golden metrics while the canary is still at 1%.

    Lands in #deploys in Slack

  • A pull request is opened

    Infra review

    List the infra changes needed before this merges, or comment LGTM.

    Lands in Comment on the pull request

  • Every 24 hours

    Memory drift

    Flag pod memory that is climbing, ignoring restarts.

    Lands in #sre-daily in Slack

  • An alert fires

    Run the database checklist

    Check replication lag, pool, slow queries, and migrations before anyone is paged.

    Lands in PagerDuty incident note

  • Every hour

    Payload size creep

    Spot a rise in large payload dumps and name the caller that changed.

    Lands in #platform in Slack

  • An inbound webhook

    Probe the endpoint

    Probe /v1/quote and check the response still comes back with a price.

    Lands in #platform in Slack

Build your own custom agent

Pick a trigger, describe the job in plain English, choose where the answer lands. Change the trigger and the jobs change with it.

Trigger (pick one)
  • Every deploy, just after rollout
  • A pull request is opened
  • Every 24 hours
  • Every hour
  • An alert fires
  • An inbound webhook
The job, in plain English
  • List the infrastructure modifications that need to be made before this merges. If there are none, comment LGTM. Otherwise share the complete list of changes in detail.
  • Flag anything in this pull request that is going to add latency or memory. Call out queries that will run without an index.
  • When this pull request adds or changes a query, check whether the tables it touches already have an index that covers it. If they do not, say so on the pull request, before it reaches production.
  • When a PR gets opened in data-service, check whether the changes can break the build. The go.mod gets bumped to a new Go version and the Dockerfiles don't, so the build fails.
  • Check whether this changes an alert threshold or a runbook step, and say which docs are now wrong.
Lands in
  • #deploys in Slack
  • PagerDuty incident note
  • Comment on the pull request
  • Linear ticket
  • #sre-daily in Slack
  • #platform in Slack
  • #incidents in Slack
  • Reply to the webhook

A pull request is opened → Flag anything in this pull request that is going to add latency or memory. Call out queries that will run without an index. → Comment on the pull request

Deploy this agentFree to start. No credit card.

These are examples, not a menu. Anything you can describe, an agent can run.

The full custom agent library

Every trigger, and the jobs teams hand it. None of these are fixed: they are starting points you rewrite in your own words.

Every deploy, just after rollout

  • Watch the canary

    Read the golden metrics and flag any anomaly on any server. Watch the canary while it is at 1% and tell me before we widen it.

    Lands in #deploys in Slack

  • Compare against the last good window

    Compare error rate and p99 against the same window before this deploy. If anything regressed, name the commit that did it.

    Lands in #deploys in Slack

  • Check the neighbours

    Check that every service depending on this one is still healthy, not just the one that shipped.

    Lands in PagerDuty incident note

A pull request is opened

  • Infra review

    List the infrastructure modifications that need to be made before this merges. If there are none, comment LGTM. Otherwise share the complete list of changes in detail.

    Lands in Comment on the pull request

  • Latency and memory

    Flag anything in this pull request that is going to add latency or memory. Call out queries that will run without an index.

    Lands in Comment on the pull request

  • Index coverage for new queries

    When this pull request adds or changes a query, check whether the tables it touches already have an index that covers it. If they do not, say so on the pull request, before it reaches production.

    Lands in Comment on the pull request

  • Build compatibility on dependency bumps

    When a PR gets opened in data-service, check whether the changes can break the build. The go.mod gets bumped to a new Go version and the Dockerfiles don't, so the build fails.

    Close to verbatim from a customer request.

    Lands in Comment on the pull request

  • Docs that go stale

    Check whether this changes an alert threshold or a runbook step, and say which docs are now wrong.

    Lands in Linear ticket

Every 24 hours

  • Memory drift

    Check memory usage for the pods and flag it if it is climbing. Ignore pod restarts, they reset the counter and are not the signal I care about.

    Restart-aware, so a nightly redeploy does not mask a real leak.

    Lands in #sre-daily in Slack

  • Yesterday, grouped by cause

    Summarize every alert that fired in the last day, grouped by root cause, and tell me which ones were noise.

    Lands in #sre-daily in Slack

  • Slow query digest

    Every morning, list the slowest statements of the last 24 hours ranked by total time, and tell me what changed about each one since yesterday.

    Ranked by total time rather than worst single run, so a frequent query outranks a rare outlier.

    Lands in #sre-daily in Slack

  • Disk headroom

    Find volumes that will cross 80% within a week at their current growth rate, and say what is filling them.

    Lands in Linear ticket

Every hour

  • Payload size creep

    Check the logs and tell me if there is any increase in large payloads being dumped. If there is, work out which caller changed.

    Lands in #platform in Slack

  • Idle resources

    Sweep the cluster for resources nobody is using: pods sitting far under their requests, PVCs with no pod attached, load balancers with no traffic. Post the ten most expensive, with what they cost per week.

    Capped at ten and ranked by cost, so the output stays short enough to act on.

    Lands in #platform in Slack

  • Queue not draining

    Watch queue depth on the payments worker. Tell me if it is growing rather than draining, and how far behind it is.

    Lands in #platform in Slack

An alert fires

  • Run the database checklist

    Before paging anyone, check replication lag, connection pool, slow query log, and recent migrations. Attach whatever you find to the alert.

    Lands in PagerDuty incident note

  • Have we seen this before

    Work out whether this is the same failure as the last three times this alert fired. If it is, say so and link them.

    Lands in #incidents in Slack

An inbound webhook

  • Probe the endpoint

    POST to /v1/quote with the sample payload and check the response comes back with a price field. If it does not, find out what changed.

    Lands in #platform in Slack

  • Assemble the customer picture

    Take the customer ID in the payload and assemble everything we know about their last failed request.

    Lands in Reply to the webhook

Questions this page usually raises

Monitoring tells you a threshold was crossed. Sherlocks AI starts from that signal and does the work that follows it: correlating recent deploys, reading logs and traces across services, checking infrastructure state, ruling out causes, and reporting what actually broke. It does not replace Datadog, Prometheus, or Sentry. It reads from them, along with your cloud and orchestration layer, and turns what they show into an answer.

Start your free month

Cancel anytime, fully refundable.

A note from the founder

You'll talk to me, not a sales rep. Bring your worst incident from last quarter.
Gaurav ToshniwalGaurav ToshniwalFounder and CEO, Sherlocks AI

Tell Me About Your Worst Incident

We use cookies

Only the strictly necessary ones are set when you arrive. Analytics and marketing cookies are set solely if you allow them, and you can change your mind at any time. See our Privacy Policy.