Required tool sequence
Confirm that lookups, policy checks, and approvals happen before the controlled action.
Tool-calling agent testing
Test AI agent tool calls before the wrong action ships.
The agent can choose the expected tool and still skip a required lookup, carry forward the wrong customer ID, call tools in the wrong order, or execute the same action twice. Konsista replays the complete action path before release.
The right tool can still produce the wrong action.
AI agent integration testing must follow the path from evidence to action. Verify what the agent checked first, which values it carried forward, when it stopped, and how many times the external action occurred.
Critical arguments
Track recipient, record ID, amount, account, destination, and other values where drift changes the real-world outcome.
Action multiplicity
Detect repeated sends, writes, refunds, charges, bookings, or deletes even when the final response appears normal.
Required and forbidden tools
Show when a required tool disappears or a prohibited action enters the candidate path.
Evidence/action consistency
Compare the observed lookup result with the action arguments so the agent cannot act against its own evidence.
Repeat stability
Run the same scenario several times to expose intermittent tool-order and argument drift hidden by one passing run.
Declare the procedure, then inspect literal evidence.
A test case describes the behavior your team already expects. Konsista does not invent the policy or use an LLM judge to decide whether the action was good.
- Provide a risky request and the tools available to the agent.
- Declare the required sequence and the arguments that must remain stable.
- Replace real side effects with a mock or sandbox tool layer.
- Run baseline and candidate configurations repeatedly.
- Review deterministic findings and the explicit coverage statement.
Example procedure evidence
lookup_account → request_approval6/6
request_approval → delete_record2/6
recordId = customer-4829stable
delete_record max count = 1held
Mock the action boundary.
Repeated tests must never send a real message, delete a real record, charge a card, submit a claim, or modify production data. Konsista observes mocked or sandboxed tool calls and reports the declared procedure. It is not a runtime permission system and does not replace production guardrails.
Related AI agent testing guides
Test the tool path you cannot afford to get wrong.
Send the action, what must happen before it, and the target that must remain correct. The first run uses synthetic records and mocked tools.
Test one tool path