Agentic Testing

Test Cases

Overview

Silverile test cases can run three ways. Manual means a person clicks through the steps by hand. Automated means a stored Playwright script, executed by a runner you host yourself. Agentic is the newest: there is no stored script at all. Instead, a coding agent you already use — Claude Code, Codex, GitHub Copilot agent mode, or Google Antigravity — reads the test case's plain-English steps and expected result, drives your actual running application itself to carry them out, and reports back exactly what it observed.

The practical difference from Automated shows up the moment your UI changes. A Playwright script only knows the exact selectors it was recorded with — a moved button or a renamed field breaks it until someone re-records it. An agentic test case has nothing to re-record: the agent re-reads your current application every time it runs and works out how to carry out the steps against whatever is actually there right now, the same way a person would.

You stay in control

Silverile itself never touches your codebase or your application — it never runs code and never has access to your repo. Your own agent does the driving, using the same tools and permissions you've already granted it for writing code, on infrastructure you control. Silverile's role is coordination and record-keeping: it tells the agent what to test and stores the result, nothing more.

When an agentic test finds a bug, your agent does not fix it unaware of your wishes. It explains what broke and asks first — see When your agent finds a bug below.

Manual vs. Automated vs. Agentic

ManualAutomatedAgentic
Who or what runs itA person, by handA stored Playwright script, on your runnerA connected AI coding agent, live
Setup requiredNoneWrite and maintain a JSON scriptNone — plain-English steps are enough
Breaks when your UI changesNoOften — the script needs re-recordingRarely — the agent works out each run fresh
Needs something connectedNoYes — your own runnerYes — your own coding agent
Can it fix what it findsN/ANoYes, only with your permission

What you need

Agentic testing runs over MCP (Model Context Protocol) — the same connection your coding agent likely already uses to read and edit files in your repo. Once it's connected to Silverile, the exact same connection also lets it read test cases, drive your running application, and record results back — no separate integration to set up. Four agents are currently supported: Claude Code, Codex, GitHub Copilot agent mode, and Google Antigravity. If yours isn't connected to Silverile yet, just ask it — "connect to Silverile's MCP server at https://mcp.silverile.com/mcp" — and it will walk you through its own exact connection steps.

Setting up an Agentic test case

1Open the test case you want to convert (or create a new one)
2Set Execution Mode to Agentic instead of Manual or Automated
3Write (or keep) its steps and expected result in plain English — there's no JSON spec to write, unlike Automated
4Save. No further configuration is needed before running it

Running one — a real example

Say you have a test case titled "User can reset their password", with steps like "Go to the login page, click Forgot Password, enter a valid email, submit" and an expected result of "A confirmation message appears and a reset email is sent." You tell your connected agent:

"Test story 42."

Here is what actually happens, step by step:

1Your agent fetches the test case from Silverile — the steps and expected result above
2If it doesn't already know where your application is running, it asks — a local dev server, a staging URL, wherever you tell it
3It drives a real browser to your login page, clicks Forgot Password, types in a test email, and submits — exactly as a person would
4It compares what actually happened against the expected result — did a confirmation message appear? Did anything look wrong?
5It records the outcome in Silverile with a real description of what it saw, e.g. "Confirmation message appeared; reset email received within 5 seconds" — not a bare pass/fail

If no confirmation message had appeared, your agent would not simply mark it failed and move on — see the next section for what happens instead.

When your agent finds a bug

Your agent does not fix anything on its own just because it found a problem while testing live. It explains what broke — what it did, what it expected, what actually happened — and asks before making any code change. Whoever is running the session might be QA, who shouldn't casually approve a code change, or the developer themselves, who might say yes immediately — either way, the decision is always yours, every time, not a one-time setting you configure and forget.

If you approve a fix, your agent re-runs the exact same live check again afterward — it never records a test as passed just because a code change looks correct. A fix is only ever recorded once it has actually been proven to work, live, the same way the original failure was.

Evidence

Visual evidence — a screenshot of what your agent actually saw — can be attached directly to the test run itself, appearing in Silverile the same way any other attachment does. This works for a pass or a fail, and whether the run was agentic, automated, or manual. You don't need to ask for this separately; your agent attaches it on its own when it has something worth showing.

If a failure gets filed as a defect, evidence can be attached there too — independently of whatever's attached to the run. The two aren't linked to each other; attaching to one doesn't attach to the other.

Troubleshooting

My agent says it can't find a running app to test against

Tell it where your application is running (a local dev server URL, or a staging environment) — it asks rather than guessing wrong. If nothing is reachable this session, it falls back to reviewing the code instead and says so explicitly.

How do I know whether a test was actually run live, or just reviewed in the code?

Your agent states this explicitly every time, in both the recorded result and its summary back to you. It never lets the two look the same.

The Execution Mode dropdown doesn't show Agentic

Confirm you're editing the test case (not just viewing it) — Execution Mode is set from the test case's edit view alongside Manual and Automated.

My agent isn't finding any Silverile tools at all

It isn't connected to Silverile's MCP server yet. Ask it directly — "connect to Silverile's MCP server at https://mcp.silverile.com/mcp" — and it will walk you through the exact steps for whichever tool you're using.

Can my agent test against production?

Technically yes if it can reach the URL, but there's no guardrail stopping it — treat this the same as you would any other live testing against production, and point it at a staging or local environment unless you specifically intend otherwise.