deltaZ builds working copies of the systems your agent uses, like Gmail, Slack, Stripe or your own API. It runs your agent through hundreds of scenarios on every change and tells you what broke, where, and how to fix it, before your customers find out.
Early access. We are working with a small group of agent teams.
A small change to a prompt, a tool description or a model can quietly break an agent. Most teams find out from users. deltaZ gives you a safe place to find out first.
Pick a system, a starting state and the feature you care about. deltaZ generates the scenarios, from the happy path to the messy cases: slow APIs, locked records, missing permissions, users who change their mind.
Example: broadcast email agent on Gmail
Run the full suite on every commit, prompt edit and model upgrade. If behavior gets worse, the merge is blocked and you see the exact step where the agent went wrong.
Works with GitHub Actions, GitLab CI, any runner
Each failure comes with a suggested change to the prompt or tool. For teams that train their own models, failed and corrected runs become training data.
For API models and open models
Every agent runs in the same worlds against the same checks, so scores mean the same thing across versions, models and vendors.
Useful for model swaps and vendor evaluation
Your agent connects to the twins the same way it connects to production: over the normal API or MCP. Nothing in your agent needs to change.
Choose a ready-made twin, or upload an openapi.yaml and a sample of traces. deltaZ builds a stateful copy with the same endpoints, rules and realistic data.
Pick a starting state and a feature. deltaZ writes the test cases and lets you add your own. Faults like timeouts and locks can be injected at any step.
Each run gets a fresh, isolated copy of the world. Hundreds of runs finish in minutes, and nothing touches your real systems.
Checks look at what the agent actually did, not just what it said. You get pass or fail, the step that failed, and a suggested fix.
This walkthrough uses sample data. It shows the kind of failure that slips past normal tests and how deltaZ surfaces it.
A billing agent applies a credit to a customer account in the company's ERP. The team edits the system prompt and ships version three.
In the twin, deltaZ locks the database record halfway through the write. The agent retries without checking what already happened, uses 180,000 tokens and posts the credit twice.
deltaZ compares v3 to v2 and finds the cause: a sentence about locked records was removed. It blocks the merge and proposes the fix. With the fix, the same 1,000 runs go from 7.4% failures to 0.3%.
All of this happens in the twin. Your real ledger is never touched.
A test is only useful if the world behaves like the real one and the check measures what matters. That is where we spend our time.
Emails stay sent, charges stay charged. Each action changes the world the way the real system would.
Timeouts, rate limits, locks, partial failures and bad data, placed at the exact step you choose.
Every run starts from a clean, known state, and runs execute in parallel.
Code checks the end state where it can, like "was the card charged once?" A model grades the rest.
See the full trace of a failed run and the step where it went wrong, not just a score.
Agents connect over plain HTTP or MCP. LangGraph, the OpenAI or Claude SDKs, or your own code.
Scenarios and checks live next to your code, so a change to the tests is reviewed like any other change.
Export failed runs with corrected trajectories, ready for fine-tuning or prompt optimization.
Before Apollo 11 left the ground, it had landed on the Moon over and over, in simulation. When an alarm fired minutes before the real landing, mission control had seen it before. Agents deserve the same: a world to practice in before the real one.
deltaZ product demo · 2:57 · numbers shown are sample data
We are onboarding a few teams who are building agents on top of real systems. Tell us what your agent does and which systems it touches, and we will set up a first suite with you.