Observe
Agents connect to your repository, CI and field data and record how each page and build performs today.
01
Yamnik · PlatformAI optimization agents
Release 2.4 · October 2026Bundle · Vitals · Reliability
Build · Flow · Growth01
Release 2.4 · October 2026Bundle · Vitals · Reliability
Build · Flow · Growth01
Yamnik: self-optimizing software
Agents that find slow code and open the fix as a pull request.
- Specialist agents
- 6
- Profiling
- 24/7
- Merge rights
- 0
Continuous optimization platform
Make every release faster than the last.
Six agents watch bundles, vitals, builds, errors, journeys and search on every commit and every deploy.
Each finding becomes a small pull request with the before-and-after numbers attached.
Playbooks proven in production at
- Wakacje.pl
- Crustlab
- Fujitsu
- Blue Owl
- Wakacje.pl
- Crustlab
- Fujitsu
- Blue Owl
How it worksThe loop
Observe · Measure · Propose · ProveRuns on every commit
and every deploy02
Observe · Measure · Propose · ProveRuns on every commit
and every deploy02
Four steps,
every deploy.
Most teams do performance work once, in a quarter somebody fought for. Yamnik repeats it on every deploy and never ships a change it has not measured.
Measure
Each problem the agents find gets a baseline in bytes, milliseconds, build seconds, incidents and euros.
02
Propose
The fix arrives as a small pull request with the evidence and a rollback path.
03
Prove
After merge, the agent measures again in production and adds the result to the pull request.
04
An optimization
team in your CI
Each agent has one specialty, and none ships a change without a measurement.
Yamnik splits profiling across six agents that share one evidence store. They cover bundle size, Core Web Vitals, reliability, build time, user flows and search growth.
Every agent began as a playbook our engineers ran by hand on production systems at Wakacje.pl, Crustlab, Fujitsu and Blue Owl. Now it runs on every commit.
Live runWatch an agent
work a repositoryPick any agent
to replay its run04
work a repositoryPick any agent
to replay its run04
Watch an agent
dig.
Every run follows the same steps: attach, baseline, scan, propose, verify. Each run ends in a pull request your team can review.
yamnik / bundle-agentacme/storefront@main00/07
- connecting to acme/storefront…
Sample run · demo repository · timings compressed
The fleetSix specialist agents
One shared evidence storeEach built from
a client engagement05
One shared evidence storeEach built from
a client engagement05
Six agents.
One fleet.
Each agent owns one layer of your product and hands evidence to the others, so a bundle fix that hurts reliability never makes it to review.
ClientsTravel · iGaming
Enterprise · E-commerceFive years
Four engagements06
Enterprise · E-commerceFive years
Four engagements06
Proven in
production
Each agent encodes an engagement our engineers ran inside a client codebase, under deadlines and alongside feature work. We measured every number before and after.
The receiptsMeasured before
and afterProduction numbers
from client work07
and afterProduction numbers
from client work07
The receipts
These are production numbers from client engagements, measured before and after each change. Every agent applies the same check to its own pull requests.
Gains estimatorWhat is slow
costing you?Conservative model
Sources below08
costing you?Conservative model
Sources below08
Price your
milliseconds
Enter four numbers you already know. The estimator shows which gains the agents would go after first, and the audit replaces the estimate with measurements.
(01) Your numbers
(02) What the fleet could find
€360K
estimated revenue upside per year, based on a 6.0% conversion lift from 1,200 ms faster pages.
- LCP after
- 3.6 → 2.4 s
- CI wait recovered
- 81 h / mo
These are estimates. We assume the fleet removes a third of your current LCP, but never below 1.8 s, and apply a 0.5% conversion lift per 100 ms saved, capped at 15%. That is about a sixteenth of the 8.4% retail uplift per 0.1 s that Deloitte measured in Milliseconds Make Millions (2020). Build savings assume 40% shorter builds and 6 builds per engineer per working day; our measured case was 55%.
PrinciplesHow the agents
are allowed to workNon-negotiable
on every plan09
are allowed to workNon-negotiable
on every plan09
(01)
Nothing ships
without a number
Every pull request carries a baseline, a projection and, after merge, its production result.
(02)
Agents propose.
Humans merge.
No agent holds merge rights. Your reviewers decide what reaches production.
Measured first,
merged by people
PlansStart with an audit
Keep what pays offPrices in EUR
excl. VAT10
Keep what pays offPrices in EUR
excl. VAT10
Next stepGains Audit
Fixed price, fixed scopeTwo weeks
One repository11
Fixed price, fixed scopeTwo weeks
One repository11
Find out what
slow costs you.
Start with a two‑week 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.