Skip to content
NewIn release 2.4, the Reliability Agent reads Sentry traces and Playwright reports
Yamnik · Product01

Profilersthat ship.

Six agents, one for each layer of your product.

Yamnik connects to your repository, CI and production telemetry. The agents profile each layer and return every gain they find as a pull request.

Measured first
Every finding has a baseline, a projection and a production result, or it does not ship.
Reviewed by you
Agents open pull requests. Your engineers decide what merges, and when.
Small changes
One gain per PR, sized for a ten-minute review and a one-click revert.
The loop02

A measured
rollout, always

  1. 01ObserveA least-privilege app connects to your repo, CI and telemetry. Agents map your routes, journeys and performance budgets.Baselines per route
  2. 02MeasureAgents profile each problem on its own and measure its cost in kilobytes, milliseconds, minutes, incidents and euros.A ranked, costed backlog
  3. 03ProposeThe agent writes the smallest change that captures the gain, with tests, notes and a rollback path.A reviewable pull request
  4. 04ProveAfter merge, the agent re-measures in production and files the result next to its projection.A receipt for finance
The fleet03

The six
agents

Each agent covers one layer and shares evidence with the others. Our engineers ran each agent's checks by hand on client work before automating them.

  1. YMK-01

    Bundle Agent

    Finds the bytes nobody needs.

    Built at Wakacje.pl · 1.8 MB → 500 KB

    Profiles every route's client payload, traces each kilobyte to the import that caused it, and removes what users never execute.

    Watches

    • Per-route JS
    • Duplicate packages
    • Serialised data
    • Dead dependencies

    Typically finds

    • Two copies of a design system shipping together
    • Server-only data serialised into the client payload
    • Heavy imports with lighter, drop-in equivalents

    Reads from

    • webpack / Turbopack / Vite stats
    • Lighthouse CI
    • Source maps
  2. YMK-02

    Vitals Agent

    Turns slow paints into pull requests.

    Built at Crustlab · TTI 7 s → 2 s

    Watches field data for every template, isolates what blocks the main thread, and rebuilds the slow path, including iframes, fonts, media and hydration.

    Watches

    • LCP · INP · TTI
    • Third-party cost
    • Render-blocking assets
    • Hydration

    Typically finds

    • Vendor UI embedded in iframes that blocks interactivity
    • Render-blocking fonts and scripts on the critical path
    • Hero media oversized for the viewport that loads it

    Reads from

    • Chrome UX Report
    • PageSpeed Insights
    • RUM beacons
    • WebPageTest
  3. YMK-03

    Reliability Agent

    Catches the failures your users already see.

    Built at Wakacje.pl · ~800 → ~500 incidents/day

    Matches your error streams against your test reports, then fixes the root cause: missing retries, version skew, flaky gates or undocumented contracts.

    Watches

    • Error streams
    • Failed requests
    • Flaky E2E
    • Release health

    Typically finds

    • Network and 5xx failures with no retry path
    • Version skew between build and runtime after deploys
    • Undocumented third-party contracts, mapped from live traffic

    Reads from

    • Sentry
    • Datadog
    • Playwright / Cypress
    • CDN logs
  4. YMK-04

    Build Agent

    Cuts minutes from every CI build.

    Built at Wakacje.pl · Cold build 180 s → 80 s

    Times every CI step, finds the transpile passes and cache misses you inherited, and migrates them to what your framework already ships.

    Watches

    • Cold & warm builds
    • CI queue time
    • Transpile pipeline
    • Cache hit rate

    Typically finds

    • Legacy Babel configs that disable the framework compiler
    • Cache keys that miss on every CI job
    • Race conditions between build plugins

    Reads from

    • GitHub Actions
    • GitLab CI
    • Build traces
  5. YMK-05

    Flow Agent

    Walks your journeys like a user and like a screen reader.

    Built at Fujitsu · Core task time −70%

    Replays critical journeys end to end, counts the steps and seconds, and proposes the shorter path with focus, labelling and announcements fixed on the way.

    Watches

    • Task time
    • Step count
    • Focus management
    • ARIA & live regions

    Typically finds

    • Multi-screen flows that collapse into one
    • Popups that trap focus and announce nothing
    • Forms that re-ask for data already in session

    Reads from

    • Session replays
    • Playwright journeys
    • axe-core
  6. YMK-06

    Growth Agent

    Finds the traffic you are not ranking for.

    Built at Blue Owl · 125,853 → ~150,000 visitors/mo

    Compares search demand with what your catalogue exposes, then builds the missing entry pages, metadata and internal links. Editors can publish them without a developer.

    Watches

    • Indexable pages
    • Content gaps
    • Metadata
    • Internal linking

    Typically finds

    • Categories with search demand and no entry page
    • Product pages sharing one meta description
    • Orphaned pages no internal link points to

    Reads from

    • Search Console
    • Sitemaps
    • Analytics
    • Headless CMS
Anatomy of a finding04

Every finding
shows its work

Each finding names the problem, the fix and the expected gain. After merge, it adds the gain the change actually produced.

YMK-1042 · Bundle AgentVerified in production

Two copies of @acme/ui ship to every route

Baseline 1.81 MB
Projected 1.07 MB
Measured 1.04 MB
Evidence · client JS by module (KB)
  • @acme/ui (copy A)418
  • @acme/ui (copy B)412
  • promotions.json296
  • react-dom132
  • date-fns74
  • app code238
packages/web/package.json
  "dependencies": {-   "@acme/ui": "^3.2.0",+   "@acme/ui": "workspace:*",    "next": "15.4.2",
DetectedPR #482 openedMerged after reviewRe-measured +24 h
  1. 1EvidenceThe trace, stats file or replay that triggered the finding, attached in full.
  2. 2BaselineThe number before any change, measured on your infrastructure.
  3. 3ProjectionWhat the agent expects the change to win, with a confidence score.
  4. 4PatchThe smallest diff that captures the gain, plus a one-click revert.
  5. 5ReceiptThe production number after merge, shown next to the projection.
Integrations05

Plugs into
your stack

The fleet reads the signals you already collect. If a tool has a stats file, a trace or an API, an agent can read it.

(01)

Source & CI

  • GitHub
  • GitLab
  • Bitbucket
  • GitHub Actions
  • GitLab CI
  • CircleCI

(02)

Telemetry

  • Sentry
  • Datadog
  • Chrome UX Report
  • PageSpeed Insights
  • WebPageTest
  • Lighthouse CI

(03)

Frameworks

  • Next.js
  • TanStack Start
  • Remix
  • Astro
  • Gatsby
  • Vite
  • webpack
  • Turbopack

(04)

Hosting

  • Vercel
  • Netlify
  • Cloudflare
  • AWS

(05)

Workflow

  • Slack
  • Linear
  • Jira

(06)

Content & search

  • Search Console
  • WordPress
  • Contentful
  • Sanity
Guardrails06

Agents on
a short leash

Security FAQ
Least privilege
Read access to code. Write access only to the branches an agent creates itself.
Runs in your CI
Agents execute on your runners, so source code never leaves your infrastructure.
No merge rights
Every change goes through your review rules, with no admin override.
Not for training
No model trains on your code, ours or anyone else's.
Full audit trail
Every read, run and pull request is logged and attributable to one agent.
Kill switch
Pause one agent or the whole fleet from Slack, the dashboard or a config flag.
Next step07

Your first PR
in two weeks.

Start with a two‑week audit.

Fixed priceYour CI, your codeBook the audit

We connect the fleet to one repository for two weeks. You get a ranked list of gains in euros and the first pull requests, already open.