Introducing the Oscilar Agent Hub

Risk teams are about to be flooded with agents.

Agents that summarize alerts, draft dispositions, tune rules, screen sanctions hits, build workflows, explain credit decisions, and answer questions in natural language.

Some will work for isolated tasks. Most will miss the harder problem.

A single-task agent bolted onto a fragmented system still sees fragmented context. It can move faster than the workflow it replaces, but it cannot reason across the full customer relationship. It does not know what fraud was found a few minutes ago, what AML escalated last month, what policy changed this morning, or why an analyst overrode the last recommendation.

It makes one silo move faster. It does not make the institution smarter.

Today we are launching the Oscilar Agent Hub: a platform of orchestrated agents for every risk decision.

Agent Hub brings more than 30 specialized agents across fraud, AML compliance, credit, onboarding, sanctions, disputes, analytics, workflow creation, and explainability onto one shared risk memory. Every agent reads from it. Every agent contributes back to it. Every agent improves through governed analyst feedback.

Last year, we laid out the thesis: agents would change risk decisioning only if they could do more than automate isolated tasks. They would need shared context, cross-domain coordination and orchestration, explainability, automatic feedback loops, and human oversight built in from the start. Agent Hub is that thesis in production.

This is what becomes possible when the foundation is already in place.

Why isolated agents are the next fragmentation problem

The pressure to adopt agents is real. Risk teams are being asked to move faster, reduce manual interventions, explain every decision cleanly, and do it without adding headcount.

The easy response is to buy or build one agent at a time. A document parsing agent here. An alert summarization agent there. Each of them looks like progress. None of them coordinate or work on unified memory.

That is how the next generation of fragmentation gets created. The first wave will have helped. The second wave will start asking why fraud, AML, and credit agents still cannot see each other’s work, and why each agent only sees a partial view of risk.

The foundation underneath

Agent Hub could not be built as a standalone layer. To be impactful in production, agents need context and data; a real foundation underneath them.

They need a unified customer risk profile: a live record of identity signals, transaction history, device and network intelligence, behavioral patterns, prior cases, policy outcomes, and analyst decisions, continuously updated and shared across the platform.

They need cross-domain decisioning, so a fraud flag can contribute to an AML alert, an onboarding decision can inform a credit decision, and a sanctions hit can carry the customer’s full history into the decision. Agents shouldn’t just automate. They should orchestrate.

They need to be trained on the institution's own SOPs, so each agent reasons from that organization's actual risk procedures and risk appetite — not a generic prompt that treats every institution the same. A generic agent answering generic prompts moves fast and misses what matters; an agent tuned to the customer's playbook and thresholds is the one a risk team can actually put into production.

They need a real evaluation framework and feedback loop, not a thin wrapper around an LLM. Most agents on the market are a prompt and an API call, with no way to measure accuracy, catch regressions, or improve from analyst decisions over time.

They need explainability built into the work, so decisions, model outputs, rule triggers, and overrides are captured as they happen, not reconstructed later when a regulator asks.

And they need a governance layer: policy boundaries, approval workflows, model risk controls, and bias monitoring that apply before recommendations hit production.

Most agentic solutions start with the agent and try to connect it to the foundation later. We built the foundation first. Agent Hub is the orchestration layer on top of it.

Driving impact with orchestrated agents 

Three things distinguish Agent Hub from other approaches we’ve seen customers explore.

Every agent starts with the same shared context. A fraud agent, AML agent, credit agent, and onboarding agent are not each guessing from partial information. They are reasoning from the same live customer view, with the institution’s full history behind the recommendation, tuned to the business’ specific risk SOPs.

Every agent contributes to the same memory. When Fraud Disputes surfaces a counterparty pattern, AML L1 Review uses it on the next alert. When Workflow Generator ships a policy change, Credit Explainability has the rationale ready. The work carries across agents instead of resetting case by case.

Every improvement stays governed. Human-augmented agents allow for genuine governance, so that when an analyst confirms or overrides an agent solution, that becomes a feedback signal. And when an audit log is required, those same inputs and agent decisions are made available for internal and external auditors to review in plain English.

This is what most agent products will struggle to copy on top of fragmented infrastructure. The feedback loop only works if the agents share the same memory. Otherwise it is isolated tuning, agent-by-agent, with no learning across the institution.

Humans own the decision

Agent Hub is built for leverage, not replacement.

A fraud flag can lock someone out of an account. An AML escalation can end a banking relationship. A credit decline decision requires sending an adverse action notice to the customer. These decisions need judgment, accountability, and a person who can explain the outcome to a regulator, an auditor, or the customer on the other end. Agents gather context, surface signals, draft recommendations, monitor performance, and document rationale. Humans confirm, adjust, override, and own the outcome.

What changes is the human workload. An AML investigator used to open an alert, pull data from three to five systems, try to build a picture of the context behind the alert, make a judgment, document the disposition, and move on. With upward of 90% of alerts being false positives, trained analysts would spend north of 30 minutes per alert confirming what they likely already know.

The Agent Hub turns this into 10 minutes of judgment, not hours of context-gathering. And it is not only the human workload that reduces. The Rule Recommendation Agent tracks KPIs and rule performance continuously and proactively recommends the threshold changes and new rules that keep pace with shifting risk, so the institution adapts as patterns move, instead of catching up long after they have.

What Agent Hub does on day one

From day one, Agent Hub gives risk teams powerful new operating modes: they can change policies using natural language without waiting on engineering, recommend new rules and thresholds to optimize KPIs, investigate alerts at scale and act on cases with full context already assembled, document the rationale as decisions happen, and more.

Today Oscilar’s agents are helping risk teams to reduce false positives by 45%, review alerts at scale 3X faster, and deploy entirely new risk policies 5X faster, giving teams leverage and insight to focus on what they do best: apply judgement to deliver optimal outcomes.

Make risk operations more strategic

• Workflow Generator Agent turns natural-language policy instructions into versioned and governed risk workflows without an engineering ticket. Teams simply type their desired risk logic  in plain English or even upload visual diagrams that outline pre-existing policies and let the agents do the rest.

• Rule Recommendation Agent monitors rule performance, recommends new rules and threshold changes, flags redundant rules, and attaches impact analysis to enable the analyst to go live with the recommended changes with confidence.

• Analytics Agent lets risk teams query data, generate dashboards, investigate trends, and generate insights without SQL.

• Test Case Generation Agent gives you confidence in changes before deploying to production, testing new policies against edge cases. 

• Create your own agent and insert your own AI logic anywhere in your risk workflows,  with no code. Call LLMs, validate data, and automate complex analysis without writing a single line of code.
  
  

Investigate with full context

• Compliance agents to drive alert resolution in minutes, not hours: driving streamlined and context-rich first-line review with agents for AML L1 Review, SAR Narratives, and CTR Filing. Build more context with agents that screen across PEPs, Sanctions, Adverse Media, and OSINT, and make more informed CDD decisions with KYC and KYB agents.  

• Fraud Disputes Agent gathers evidence, reasons through dispute claims, and recommends next action.
  
  

Document and defend the decision

• Credit Explainability Agent generates regulator-ready rationale for credit decisions and policy changes.

• Credit Memo Agent enables defensible decision narratives without the headache, analyzing case data, enrichment, and workflow outcomes to produce a structured PDF with risk factors, decision rationale, cash flow and business insights.

Policy changes that once required a multi-week engineering cycle can be configured in minutes and routed through governance before release.

Trusted by the world’s leading financial innovators 

SoFi uses Oscilar’s Workflow Agents to deploy new credit risk strategies 50% faster, moving from weeks to days, with a 30%+ improvement in decisioning speed.

"Workflow Agents support SoFi’s ability to make swift and accurate risk decisions and serve our members’ needs. We can easily launch and iterate new policies and adapt with unprecedented speed" —Adam Colclasure, Senior Director of Risk Data and Decisioning at SoFi

"Why not build this ourselves?"

It is the obvious question. Most institutions have engineering teams. Many have tested LLM APIs. The first prototype is not the problem.

Production is the problem.

A working POC proves the model can do the job once. A production-ready agent must detect performance drifts, explain where each recommendation came from, outline what humans overrode, flag what changed in the data, and ensure that the institution can reconstruct and explain the decision later.

None of those core capabilities ship with such simplistic "LLM wrapper" agents.

Drift detection, an evaluation framework, a feedback loop, override tracking, full decision reconstruction — all of it has to be built, owned, and maintained. And a missed sanctioned entity does not go in a Jira ticket. It goes in an exam report.

Maintenance is permanent. Engineers move on, and the agent they built becomes a black box to everyone else. Every new agent on a build path is another three months of engineering plus another permanent operational commitment. PEP today, EDD in six months, KYB after that. On Oscilar, each new agent builds on the same platform and becomes a configuration change via a self-service UI.

The first agent is easy to prototype. The second one means you are building an entire agent orchestration platform. We already built it.

What changes now

Risk teams do not need a larger collection of disconnected agents. They need agents that work together: reading from the same customer context, contributing back to the same memory, improving through governed feedback, and supporting decisions a human can own.

That is what Agent Hub delivers: more than 30 specialized agents across fraud, AML, credit, onboarding, sanctions, disputes, workflow, analytics, and explainability, all operating on the foundation built to support them. Deploy specialized agents from the Oscilar platform in minutes, or integrate Oscilar’s Agent Hub on top of your existing risk solutions.

More than 100 institutions trust Oscilar to process tens of billions of automated risk decisions a year, each in under 100ms. Today, we are giving their teams the orchestrated agent layer that does what isolated agents can't: make the entire institution smarter, not just faster.

Learn more about Agent Hub

NEW GUIDE

Agents in AML: A Production Playbook from the Investigator's Seat

Get your copy

->