Application journey

Every agent that moved this application — voice and process — the decision it made and why, laid on the pipeline. Each card says which store the claim came from, so nothing here is unattributable. A stage with no recorded decision says so rather than being filled in.

Calls

Every call the agents have taken on this account, newest first — read from the platform’s own event bus, not a mock. A call that moved the application forward says so on its row, with the reason it was allowed to move.

AutoApprove agent fleet

Every agent working this account. The process agents are read live from the platform, so this is the deployed inventory rather than a diagram; the voice agents are the four that answer the phone. To an operator they are one team — whether a job runs on a call or as a workflow is an implementation detail.

Voice seam

The seam’s per-agent view of this borrower, read live from the record on every refresh. Each agent sees only its own scoped fields; absent facts stay absent, because a placeholder is a fact an agent will read out.

⚠ Not yet identical to what the live voice agent receives. This page projects the seam’s own field table; the phone path builds its context separately and requests a different set. They overlap but diverge — for the servicing agent, 2 of the 5 fields it requests. Until the two are reconciled, read this as the seam’s view, not as proof of what a caller was told. Tracked in LFP-0031.

My queue

Items waiting on you — decisions the agents have routed to a person in your role.

Deals

The full book — search, filter, and sort every application the platform has processed.

Agent flows

The live workflow graph behind each stage — pulled straight from the running agents, not a mockup.

Decision trace

Every agent run this account has executed, newest first — read live from /v1/traces. Open one to see the exact path it took: which router, gate, knowledge and tool nodes fired, in order, each with the time it took. The mechanism behind a decision, not a summary of it.

Evidence

Enter an application and read its decision file — what the seam knew, how identity was checked, who moved the file, and which store each claim came from. Read live from the grounding record and the journey; a step with no record says so.

Evidence v2 — governance & decision trace

The loan origination→disbursement journey as an action ledger — who acted (deep agent vs manual operator) under what authority, which SOP it adhered to, and the immutable trace behind every step. Read live from the grounding record, the run trace, and the audit chain.

Governance & audit

Every agent run this console has executed — the human authorizations, the outcomes, and a link to each run's node-level trace.

Agent evaluations

How well the agents perform, scored by an automated LLM-judge on every run — accuracy, groundedness, latency and cost, per stage and over time.

Human review

Human-in-the-loop quality: consultants rate the agents' decisions, and we track agreement, overrides, and the review backlog — the ground truth behind the automated scores.

ROI & business impact

What the agents are worth — cost avoided, capacity unlocked, speed to fund, and margin. The CFO case.

Compliance & model governance

Defensible, explainable, fair — adverse-action reasons, fair-lending monitoring, SR 11-7 model governance, guardrails, and one-click audit packs.

Agent reliability & AgentOps

Production rigor — SLOs, drift monitoring, versioning & safe deploys, eval suites, cost/FinOps, and the human-feedback flywheel.

Lender & borrower intelligence

NOVA marketplace performance and the borrower journey — who approves, who prices best, and where borrowers drop off.

Settings & admin

Users and roles, agent configuration, SLA thresholds, integrations, and workspace white-label.