Try QuadRails
Prove a guardrail works before you depend on it.
Run the same scenario twice, once unprotected in an isolated sandbox and once with QuadRails on. Compare the outcome, the trace, the latency and the cost, side by side.
Test scenarios
Test before you trust.
Run your multi-agent system against the same scenario twice, with guardrails off in an isolated sandbox and on in a protected run. Compare the outcome, the trace and the cost.
Unapproved tool call
User asks for an account summary. The agent proposes delete_records.
Guardrails off (sandbox)
Tool executes. 412 synthetic records removed.
No record of why the call was allowed.
Guardrails on
Blocked before the backend. Zero calls made.
Reason: tool not in the agent's manifest for this task.
Injection inside a document
A retrieved file tells the agent to export customer data to an outside address.
Guardrails off (sandbox)
Agent follows the hidden instruction and sends the export.
The retrieved text was treated as trusted.
Guardrails on
Source marked untrusted. Export blocked.
The trace shows exactly where the instruction came from.
Ambiguous “clean up” request
User asks to clean up old records. The agent proposes an irreversible delete.
Guardrails off (sandbox)
Records deleted on the agent's reading of “clean up”.
Nobody was asked.
Guardrails on
Paused for a person to approve, edit or reject.
The approver sees the evidence, not the agent's summary of it.
Tool use the task allows
User asks for an account balance. The agent proposes get_balance.
Guardrails off (sandbox)
Balance returned.
Same answer, no evidence trail.
Guardrails on
Allowed in 14 ms. Balance returned.
Not everything gets blocked: in-scope work runs at full speed.
Team budget exceeded
An agent loops on repeated model calls past the team's daily cap.
Guardrails off (sandbox)
Calls continue until someone notices the bill.
Spend is reconciled weeks later.
Guardrails on
Reservation fails at the cap. Agent throttled.
Routed to a cheaper approved model or stopped, per policy.
How a session runs
Twelve minutes from intent to evidence.
Every session uses synthetic data and an isolated environment. Guardrails are never switched off on real workloads.
Set the scene
The user's intent, the agent's identity, its tool manifest and the risk policy.
Block
An unapproved tool is stopped, and the backend call count stays at zero.
Allow and review
In-scope work runs at full speed; ambiguous work waits for a person.
Trace an injection
Follow a hidden instruction from a retrieved document to the point it's caught.
Catch drift
A new agent version loses grounding, and the release gate stops it.
Cap the spend
A looping agent hits its team budget and is throttled.
Agent readiness
Completeness alone is not readiness.
We benchmark agents on expert-written scenarios with golden answers, run each one many times, and score correctness, completeness, consistency and hallucination. Reliability across repeat runs is what gates a release.
6,248
scenario runs in a recent enterprise agent benchmark
344
expert-written questions with golden answers
85%
completeness, the easy part
53%
consistency across repeat runs, the part that gates release
Judging the judges
A cheaper judge has to earn its place.
When a small model checks whether a tool call fits the user's intent, we test it against a human-labelled set instead of trusting a rule of thumb.
False allows
Risky actions the judge let through. The number that matters most for security.
False blocks
Legitimate work the judge stopped. The number your users feel.
Review coverage
How often the judge says "not sure" and hands off to a person.
Calibration
Whether the judge's confidence matches how often it is right.
Latency
p50 and p95 per check, with timeouts tracked separately.
Cost per 1,000 checks
So the guardrail's price is weighed against the risk it removes.
Bring one agent. Leave with evidence.
We'll run your scenarios with guardrails off and on, and hand you the traces.