CLOSE THE LOOP / THE AGENT ARENA
Make your move.
Watch the world react.
Two agents. Synthetic tools.
Consequences that persist.
Test an agent’s trust boundary, choose a control, and see the consequences unfold. The challenge isn’t over when something breaks.
A red team challenge
with a second act.
Find where trust breaks.
Change the boundary.
See what happens next.
CHOOSE YOUR WORLD / 001
Every agent has
a pressure point.
Start with one of two business workflows. Each campaign unfolds over three rounds, with a new condition and a decision to make.
THE AGENT EXPERIMENT / 002
SANDBOXED CONSEQUENCESAtlas/your next decision
Test a customer operations agent. Protect the refund policy while preserving its ability to help eligible customers.
What happens in hosted AI mode?
When you run a live round, your synthetic message and campaign history are sent through the arena’s hosted service to the provider named on your selected target. Credentials stay on the server. No model weights are downloaded to your browser.
Practice runs entirely in this sandbox using the explicit action JSON. Live targets become selectable only when the service reports them available. A failed live request never silently becomes a practice result. Use synthetic content only; do not enter sensitive information.
Choose who makes the next move.
Availability is reported by the hosted service.
Same state, same input, same control. Each selected target gets an independent copy of this round.
Select at least two available targets. Comparisons do not advance your main campaign.
A system in motion
WORLD STATEINPUT
WORLD
Trusted context
Policy boundary
Your intervention
Introduce an instruction through the untrusted input. See what the agent proposes, and what the system actually permits.
Synthetic data only. Your message reaches the agent through an untrusted surface.
THEN OBSERVE.
Ready for your first move.
The world is unchanged.
Run a round to see what your intervention sets in motion.The response
The agent’s response will appear here after you run a round.
A proposed action is not an executed action.
The action trail
Boundary checks and executed actions appear here.The difference is where governance becomes real.
Your campaign, round by round.
A TRACE, NOT JUST A SCOREA first move.Find the failure.
→02A new boundary.Choose your control.
→03A stronger claim.Retest what matters.
What did you
actually change?
The debrief unlocks after three rounds.
Connect the agent’s behavior to the controls you chose and the world you left behind.TAKE THE EVIDENCE WITH YOU / 003
Make the finding
useful to someone.
Keep your prompts, decisions, and observed consequences in one evidence trail. Take it into GitHub for review, remediation, and the next test.
Download the GitHub-ready kit
Opens a prefilled GitHub issue for your review.
Nothing is submitted automatically.
FROM THIS WORLD TO YOURS
The next agent
is your system.
A sandbox makes the lesson visible. Syntony brings evaluation, governance, and assurance to the systems your business depends on.
Scope an AI assurance assessment