AgentMarketplace fraud v3
DraftEval +0.0%
- Policy
- GuardrailsBlock region: EU · Require human review > $1,500
- Eval set1,284 cases · 92% precision · 88% recall
- Precision0.0%+0.0
- Recall0.0%+0.0
- p95 latency0 ms−0 ms
Custom agents
Custom agents for policy decisions at scale
Build an agent around the exact policy decision your team needs to operate: prompts, outputs, chats, streams, reports, appeals, or high-risk escalations. Cinder agents are trained on your policies, examples, and reviewer decisions, so clear cases move fast and hard calls escalate with context.
Overview
Your policy, encoded into the operating loop
Generic classifiers can catch generic violations. They cannot understand the product context, policy nuance, and risk thresholds that make your decisions defensible.
Cinder custom agents turn your policies, examples, evals, and reviewer decisions into workflows your team can inspect, measure, and improve. Use them for GenAI safety, UGC moderation, live surfaces, appeals, policy QA, and sensitive escalations without forcing every problem into a generic classifier.
Capabilities
How custom agents work
Configure on top of your policies
Define the policy, the data the agent should consider, and the actions it can take. No model training degree required.
Live training on your team's reviews
Every reviewer's decision becomes training signal. The agent gets sharper at your problem, not someone else's.
External tool use
Agents can browse the web, hit your APIs, and pull third-party signals when the call requires more context than any single input provides.
Backtesting before production
Run new agents against historical data to see what would have changed: false positives, false negatives, queue impact, before anything ships.
Co-built when you want it
Spin one up yourself, or partner with our services team to design and train it alongside your operators.
Capabilities
What teams have built
01
Prompt & output safety agents
For validating prompts, model outputs, jailbreak attempts, and generated content before and after launch.
02
Live surface moderation agents
For streams, comments, chats, uploads, and high-volume UGC decisions that need fast, policy-grounded enforcement.
03
Child-safety signal agents
For extracting and escalating sensitive child-safety signals while keeping human review in control.
04
Policy QA agents
For checking reviewer decisions, agent performance, and policy drift across high-risk queues.
05
Appeals and escalation agents
For routing hard calls with case history, policy context, and evidence attached.
“What Synthesia needs now is infrastructure that can keep pace with how quickly the product and the threat landscape are moving, which is why Synthesia is combining its know-how with Cinder's technological capabilities.”
Synthesia