North star: make every TE customer a 1000x lender. Agents work the whole book, context makes every action right, and
each lender's intelligence compounds into AI that works only for them.
We get there in five layers. Customers meet us wherever they work (Interface). Agents do the work and earn the revenue
that funds everything else (Agentic). We build it all at software speed (Autonomous development). TE's data is what
makes the agents worth using (Context). And every run makes the next one better, per lender (Intelligence).
FIG.1 — 5 layers · click any blockGAThird-party GAFeature flag & pilotReadyNot started
Loop every agent emits traces → data collection → evals · prompt composition · routing · post-training → better, cheaper models that work only for each lender
01 · Interface
Interface
Wherever it works for you, that's where we meet.
How customers and their agents get into TE: optionality, not one door. Three classic interfaces customers already use,
and two modern ones: our agents through TE Assistant, or your agents through MCP.
01 · Interface · Classic
Classic
The interfaces TE customers already know. All three are in production; each gets uplifted so agents show up inside it.
01 · Interface · Classic · UI
UI
the TE CRM
The TE CRM your teams already know and work in every day. Agents appear where the work already happens.
01 · Interface · Classic · API
API
TE public APIs
TE's public APIs, so developers can build on TE.
01 · Interface · Classic · Mobile
Mobile
voice · SMS · one click
The TE mobile app, where we start exposing agents on the go: by voice, by SMS, with one click (approve, send,
book).
01 · Interface · Modern
Modern
The two agentic doors. Every agentic use case is exposed through both.
The one entry point to TE's intelligent platform. It's an agentic feature, but it lives in the Interface layer because
it's the common interface for every TE customer: TE agents, partner agents and your own custom-built agents are all
reached from one assistant, instead of hunting for ten features.
For customers who want to consume the TE AI platform from their own agent. Every TE capability, sub-agent and tool is
exposed over MCP, so the LO's own ChatGPT or Claude, and the lender's own agents, can work inside TE: signed in as the
LO, with contact changes drafted for the LO to confirm.
Where we focus right now. These are the agents that do the actual work, turning customer moments into appointments,
applications and funded loans. The revenue they drive funds every other layer. Agents come from three places: TE builds
them, partners build them, or lenders build their own. All of them are reachable from TE Assistant.
02 · Agentic · TE in-house
TE in-house
Agents the TE team builds. Scenario, SMS and Agentic PPE work for users; Journey and Content are for admins, though they
do a lot more than that. TE Assistant is built so any team inside TE can ship an agent and have it show up there.
An agent that builds the lender's pricing engine for them from its rate sheet. AI maps the sheet; every quote is then
calculated straight from the sheet and traced to its cell.
Agents built outside TE. They use TE data and are exposed through TE Assistant and MCP, next to TE's own agents.
02 · Agentic · Partners & third party · Sales Assistant
Sales Assistant
Sela
Third-party GA
Voice agent from our partner Sela. It calls leads, qualifies them, and warm-transfers them to the LO, available through
TE Assistant.
02 · Agentic · Partners & third party · Customer-built agents
Customer-built agents
bring your own agent
Agents a customer has already built elsewhere (on OpenAI Agents,
Claude Managed Agents,
Dify or their own stack) plug into TE's MCP and TE
Assistant, and become available inside TE.
02 · Agentic · Build your own
Build your own
We can't build every agent a customer wants. Instead, we give them the option to build their own.
02 · Agentic · Build your own · Customer Agent Builder
Customer Agent Builder
TE infrastructure · data controls · guardrails
Customers define their own agents on the TE infrastructure stack, with TE data controls and guardrails built in. The
agent is theirs; it stays within our walls. Inspired by
Dify's agent builder.
03 · Autonomous development
Autonomous development
Whatever we build, we deliver at software speed.
Headcount is no longer the constraint. The only questions are: do we need this, and can we run a pilot on it? This
layer is what lets us build at this scale and speed.
03 · Autonomous development · Coding harness
Coding harness
best-in-class harnesses · best-in-class models
We combine and use the best coding harnesses (Claude Code, Codex, OpenCode) and the best coding models. The repo carries
the instructions, skills and hooks every agent uses.
03 · Autonomous development · Background agent infrastructure
Background agent infrastructure
nobody at the keyboard
Infrastructure for agents that work end to end with nobody at the keyboard: sandboxes, credentials, schedules,
budgets and logs.
03 · Autonomous development · Self-driving codebases
Self-driving codebases
self-healing · self-supporting · self-managing
Codebases that heal, support and manage themselves: agents test, triage, evaluate, refactor and ship on their own, every
night.
04 · Context
Context
Where the actual value comes from.
The Agentic layer is what we sell. Context is why those agents are useful: we give them the lender's data. Two jobs:
Make every TE integration and data source usable by every agent in the Agentic layer, efficiently, and let agents
update it, not just read it.
Collect every data point and trace from our agents, so they can be improved, tuned and made customer-specific.
04 · Context · Agent interface + ontology
Agent interface + ontology
One interface for agents to read and update all of TE's data, plus a shared ontology (borrower, loan, lead,
journey, consent, LO) so every agent speaks the same language.
TE's journey orchestration engine. It listens for customer moments (a rate drop, an equity milestone, a new lead, a
life event) and triggers the right agent at the right time, across every channel.
04 · Context · Agent interface + ontology · LOS integration for agents
LOS integration for agents
Agents read loan status and pipeline from the loan origination system, plus the funded loan: the outcome that tells us
whether an agent actually worked.
The killer point. Collect all data and every trace from every agentic use case.
A trace is everything about one run: what the agent saw, what it did, what the LO changed, and what happened next (booked?
applied? funded?). Every block in the Intelligence layer runs on traces. No traces means no evals, no tuning and no custom
models.
05 · Intelligence
Intelligence
AI that works only for each lender.
With the Agentic layer and the Context layer in place, we can tune every agent per customer: measure it, tune it,
route it, and move it up the model ladder until each lender's intelligence is theirs alone.
05 · Intelligence · Evaluate
Evaluate
How we know an agent is good, for the industry and for each tenant.
What Harvey built for legal (Legal Agent Benchmark),
built for lending. The standard way to evaluate LLMs and agents for the lending industry: when OpenAI, Anthropic,
Google or Meta release a new model, they publish their usual benchmarks, and for lending they use ours.
05 · Intelligence · Evaluate · TE evaluation
TE evaluation
per agent · per use case · per tenant
Ready
Evaluation for every TE agent: checks on real cases, graded the way lenders grade, run before every release.
Ready today for the SMS Assistant, Scenario Assistant, Journey Assistant, TE Assistant and Content Assistant.
05 · Intelligence · Tune
Tune
Once the right data is collected, the same agent adapts to each tenant.
05 · Intelligence · Tune · Prompt composition per tenant
Prompt composition per tenant
GEPA, per tenant
The agent stays the same; its prompts are composed and tuned automatically (GEPA) to each lender's products, voice and
rules.
05 · Intelligence · Tune · Model routing
Model routing
the right model for every request
Send each request to the right model for quality, speed and cost, as
vLLM Semantic Router does.
05 · Intelligence · LLMs
LLMs
Custom intelligence, part one: the LLMs behind lending agents. A use case moves up a rung only when the next model passes
the same evals at lower cost.
05 · Intelligence · LLMs · Frontier LLM
Frontier LLM
Every use case starts on frontier models, open- and closed-weight, through AWS Bedrock, Together AI and the like.
05 · Intelligence · LLMs · Open LLM
Open LLM
Self-host open-weight models once they pass the same evals, at a fraction of the cost.
05 · Intelligence · LLMs · Post-trained LLM per customer
Post-trained LLM per customer
Post-train a model on each customer's own data with Together, River or
Modal. It ships only when it beats the frontier model on that customer's evals. That customer's
intelligence becomes theirs alone.
05 · Intelligence · Decision models
Decision models
Custom intelligence, part two: the same progression for models that score, classify, rank and route: lead score,
recapture likelihood, intent.