Mocks & Synthetic Data
Fast and cheap, but fixtures and local samples can't tell you how the real system behaves.
Free Technical Session
How monday.com Runs Agent Evals Against Real Dependencies
A new model or a prompt tweak can quietly break what your customers rely on. See how monday.com runs every PR through evals against a staging cluster with real services and data.
Featuring Dor Cohen (AI Engineering Director, monday.com) and Eyal Bukchin (CTO, MetalBear)
Want this running against your own cluster? Book a live technical session with our engineers.Bring your eval suite. We help wire it to real services in your cluster, using mirrord.
Book a technical sessionWhere evals fall behind
Mocks vs ephemeral environments vs real dependencies
A lot of what an agentic product does is call other systems: databases, internal services, third-party APIs. One changed prompt, model, or tool can quietly break those calls. Evals catch that, but only if the dependencies they run against behave like production.
Ephemeral Environments
Real behavior, but you pay for every run in spin-up time and money.
Real Dependencies
Your eval suite, run against a staging cluster with real services and data, on every PR. No environment to spin up or maintain.
What's in the recording
How to run agent evals against real dependencies
Chapters
- 09:37What is an agent evaluation?
- 13:38Mocks vs. real dependencies (staging environments)
What a suite catches once it talks to a populated staging cluster.
- 19:58What is mirrord?
mirrord lets developers and coding agents run local processes with access to a real Kubernetes cluster, without deploying first.
- 25:45Demo: running agent evals with mirrord
Wiring an eval gate to a cluster from scratch.
- 30:50Demo: how monday.com evaluates agents in development
monday.com's own framework for running agent evals.
You'll walk away with:
- The full path: eval suite → mirrord → staging services and data → merge gate in CI.
- The demo repo and a setup guide, sent to your inbox, ready to run on your own cluster.
Live technical session
Wire your agent evals to real services, with our engineers
We're offering webinar attendees a live technical session: our engineers sit with your team and help wire your agent evals to real services in your cluster, using mirrord. Here's how it works:
Watch the recording
See the eval setup monday.com runs on every PR.
Book a session and bring your eval suite
Tell us what your evals call: databases, queues, APIs.
We wire it together
Our engineers will help point your evals at your real cluster with mirrord.
Who it's for
Who this recording is for
AI & ML Engineers
You own the agent and want evals that reflect what it does in production.
PMs & Eval Owners
You own the eval suite and need scores you can trust before a merge.
Platform, DevEx & Backend Engineers
You own the cluster and want evals to use it without duplicating it.
Engineering Leaders
You want agent PRs held to the same bar as human PRs.
Who you'll hear from
Meet the Speakers
Dor is AI Engineering Director at monday.com and speaks publicly about agent evaluation, including at LangChain's Interrupt in London. At monday.com, agent PRs move through the same repository, CI pipeline, and engineering standards as human-written work.
Eyal co-founded MetalBear and leads engineering on mirrord. mirrord lets developers and coding agents run local processes with access to a real Kubernetes cluster, without deploying first. Eyal hosts the conversation and demos how to set up evals with mirrord.
Questions
FAQ
How do I book the technical session?
Pick a time here. Bring your eval suite, or just tell us what your evals call today. We do the wiring with you on the call.
How do I get the demo repo and setup guide?
They land in your inbox after you unlock the recording. If you registered for the live session, check the follow-up email. You should already have them.
Is this only for Kubernetes teams?
The demo is Kubernetes, and so is mirrord. The same problem shows up anywhere local code depends on databases, queues or APIs that only exist in a remote environment, so book a session and tell us what you're running.
Can we eval an agent against staging without deploying it first?
Yes. The eval runs on your machine or in CI while the agent talks to real databases, queues and services in the cluster. Nothing gets deployed to get a score.
Do we need a separate environment for this?
No. Several engineers and agents can run against the same cluster at once without colliding, so there's no parallel staging to build or pay for.
We already run evals with LangSmith, Braintrust, etc. Why watch?
This session will improve the results you get in LangSmith, Braintrust and other platforms you use. You'll see how to point the same suite at a real database and real services, and how to wire that run into CI.
Do I need mirrord installed to follow along?
No. Watch first, then use the repo and setup guide to run the same checks on your own cluster. Or get help while setting it up by booking a session with us.
Free Technical Session