Plate I — Specimen OverviewAI Agent Showcase · San Antonio 2026
Fig. 1
Anatomy of a Customer-Facing AI Agent
A working specimen, dissected. Seven organs that turn a chat-prompt into a production system handling sales and support at scale — and the failure mode each one prevents.
The chat panel below is alive. The screen beside this page shows its nervous system in real time.
Live Specimen
Plate II · In situ
Chat with the agent
Or reach the agent another way
The same agent answers SMS and email through the same pipeline as web chat. One conversation, one memory, regardless of channel.
Customers don't show up where you want them to. They text, they email, they fill out forms, they reply to the wrong thread. The agent has to listen on every channel a customer might use.
Failure
A chatbot that only lives on one webpage. The leads that count never see it.
Solution
A unified contact + contact-method model. One person can carry several phones, emails, and chat sessions and they all merge into a single conversation history.
IIReflexes
Classifier & router
reflexus — before the brain wakes up
Most inbound messages don't need full reasoning. A small specialized model decides what kind of message this is — new lead, existing client, scheduling reply, unsubscribe, spam — and routes it to the right downstream pipeline.
Failure
One mega-agent trying to handle every kind of message. Slow, expensive, and brittle.
Solution
Per-message classification with cheap models, plus short-circuit logic for the obvious cases that never need an LLM at all.
IIIBrain
LLM with a dynamic prompt
cerebrum — the reasoner
The model itself is a commodity. What you feed it isn't. Each turn assembles a fresh system prompt from snippets selected for this client, this situation, this channel.
Failure
One frozen mega-prompt that grows over time as new edge cases pile on. The model gets dumber, the bill gets bigger.
Solution
Dynamic prompt injection — static, data-driven, and AI-generated snippets composed per turn from a snippet library, with the rest of the codebase blind to which is which.
IVMemory
Bounded history + structured state
memoria — what survives between turns
An agent has two kinds of memory: the rolling conversation, and the durable facts about the customer. They have very different shapes and very different lifespans.
Failure
Either amnesia (no continuity across messages) or context overflow (everything dumped in, model gets dumber).
Solution
Last-N message window for the conversation, plus a structured "opportunity state" persisted separately. Client, staff, and system messages are mixed in deliberately so the agent always knows who said what.
VKnowledge
Retrieval & curated vocabularies
scientia — what the agent looks up
Reference data — product catalogs, policies, neighborhoods, options — is too large and too volatile to live in the prompt. It belongs behind a retrieval layer the agent can query.
Failure
The agent confidently invents facts. Or worse, you stuff the catalog into the prompt and watch token costs explode.
Solution
Knowledge-article retrieval for prose answers, plus curated controlled vocabularies (amenities, neighborhoods, statuses) so the agent picks from real options rather than hallucinating new ones.
VIHands
Tools, tasks, and actions
manus — what the agent can do
Talking is the easy part. The interesting agent reaches into the world: searches inventory, drafts an outbound, sends multiple messages, schedules a tour, hands work to a human teammate.
Failure
An agent that talks but can't act. Or one that acts without leaving a trail.
Solution
Tools that can emit multiple outbound messages, tasks that hand structured work to staff, and actions that update opportunity state — every call recorded so we can replay, audit, and evaluate.
VIINerves
Observability & human review
nervi — the signal humans see
The reason most production AI projects fail isn't the model — it's that nobody can see what the model is doing. An agent without observability is a liability.
Failure
Black box. No way to debug, no way to trust, no way to improve. Quality drifts and nobody notices until customers complain.
Solution
A conversation inspector that captures every thought, tool call, snippet, and context window for every message. Staff review flagged conversations daily, and those reviews feed back into an automated evaluation suite that re-runs scenarios on every change.
About this work
We build customer-facing AI agents that scale sales and customer-support staff. The seven organs above are the framework we deploy each time — the failure modes are real, and so are the solutions. Each domain we drop into is mostly the same problem in a new costume.
The agent on this page is a real production system handling apartment locating. The same anatomy works wherever consistent top performance is the bottleneck — sales, support, intake, scheduling, claims. People who can deliver it are hard to find, train, and keep.
If that sounds like your team, come find me at the table.
