Resources
Everything your teams need to evaluate an agent.
Guides for operations, risk, security and IT; the documentation we share during evaluation; and a plain-English glossary of the terms on this site.
Guides
Start here.
How it works
What ships with every agent, how it fits your stack and the five phases to production.
Security
Deployment model, AI controls, access, data handling, audit and supply chain.
The evaluation ladder
What L1 to L5 test, and when each level runs - from tool calls to business outcomes.
Path to production
Explore, Connect, Validate, Pilot, Scale - who is involved and what you come away with.
MCP servers
The connectors agents use to reach your LOS, core, bureaus, documents and payments.
Controls built in
Approval gates, deterministic engines, least-privilege tools and a full audit trail.
Documentation
The evaluation pack, on request.
Detailed documents are shared under NDA once we understand your use case, so you receive the versions that match the agents you are evaluating.
Request documentation- Architecture & deployment guide
Reference architecture, container topology, network flows and sizing for your cloud or data centre.
- Security pack
Controls, data flows, and threat model for your security review.
- Model-risk documentation
Per-agent model documentation structured for SR 11-7 style model-risk review.
- Test suite & evaluation report
The agent’s L1-L5 cases and results, ready to re-run on your own data.
- Regulatory mapping
The rules each agent touches and the guardrail or control that addresses each one.
Browse by regulation
Find agents by the rules they work under.
Each agent lists the regulations it touches. Pick one to see every agent mapped to it.
Browse by line of business
Agents for every part of finance.
Glossary
The terms, in plain English.
- Agent
- Software that carries out a defined piece of work - such as preparing a loan file - by calling tools, following a documented workflow and handing over to people at set points.
- MCP server
- A connector built on the Model Context Protocol, an open standard. It exposes a system’s functions to agents as tools with declared inputs, outputs and access levels.
- Tool
- A single action an agent can take, such as “pull credit report” or “create application”. Each tool is declared as read, write or both.
- Guardrail
- A written rule the agent must never break, for example never quoting a rate that the pricing engine did not return.
- Human in the loop
- A step where a named person reviews and decides. Regulated decisions - credit, claims, consumer contact, advice - always have one.
- Deterministic engine
- Your rules, pricing or scoring system. It produces the same answer for the same inputs, so regulated outcomes never depend on a model.
- Handoff
- The point where an agent passes work - with full context - to another agent or a person.
- Evaluation ladder (L1-L5)
- Five levels of testing: tool use, accuracy, end-to-end tasks, compliance under pressure and business outcomes. See the ladder.
- Shadow mode
- Running an agent alongside your team on live work without acting on its output, to compare results before a pilot.
- Audit trail
- A timestamped record of every tool call, rule version, model output and human decision in a case.
- SBOM
- Software bill of materials - the list of components and versions inside a release, used by security teams to track vulnerabilities.
- Model risk management
- The governance banks apply to models (for example under SR 11-7): documentation, validation, monitoring and change control.
Have a question we didn’t answer?
Tell us your use case and we’ll send what your teams need.