The architectures
AI agents built for real work.
We build AI agents in a range of shapes — and the memory layers behind them — from the simplest thing that works to the most controlled. The list below climbs from one to the other.
How the agent runs
Where it lives changes everything.
Choosing the right model is part of the job, and we help with that. But the bigger choice is structural: how much of your data the agent sees — and how it's handled before it gets there, since data often has to be cleaned, transformed, or scrubbed of personal details first — and whose computer it runs on. Each rung trades a little simplicity for a little more control.
- 1
A workflow you can see
The simplest thing that works. A tool like n8n lays the steps out as a flowchart — a trigger, the actions, an AI step wherever judgment helps — and you watch the work move through it. Fast to stand up; you can point at exactly what happens, and when.
Sounds like you if: you're a small team or an agency automating a known, repeatable process — onboarding steps, invoice routing, content production — and you want to see every move it makes rather than trust a black box.
- 2
An agent in a vendor's cloud
The agent lives outside your app and runs on a managed platform — Anthropic, OpenAI, or Gemini run the loop for you. That decoupling is what lets it act with more judgment than a fixed flowchart, and lets you swap models or scale without rebuilding. The least infrastructure to run.
Sounds like you if: you want results fast, the data isn't especially sensitive — marketing copy, public content, internal drafts — and you'd rather not run servers. A startup or a lean team shipping quickly.
- 3
Your own agent, in your cloud
The same kind of agent, running inside a Kubernetes cluster you control. Your environment, your network rules — nothing leaves your boundary except the one model call you choose to make. More to run, and far more control in return.
Sounds like you if: you're in healthcare, finance, or legal — somewhere data residency and “where does this actually run” are real questions — and “trust us” isn't good enough. The data stays inside your boundary.
- 4
A governed system in two rooms
The most controlled form. The agent runs blind, in an isolated sandbox — no keys, no database, no path out. A trusted broker in the next room holds the credentials, decides what the agent is allowed to know, and keeps the log. The agent only ever receives an approved packet and returns a result. It can't leak what it was never given.
Sounds like you if: you're handling regulated or customer data under something like HIPAA, GDPR, or SOC 2, and an audit is going to ask exactly what the agent could see — and expect you to prove it.
A note on the model itself. Every form above can call a frontier model over an API, and for most businesses that's the right choice. But the model can be private too — a privately hosted model, or a local open-weight LLM on your own hardware. Teams reach for that for one of two reasons: cost, when an agent runs a high volume of relatively simple tasks and per-call API pricing stops making sense; or data security, when regulation or policy means the data can't leave your network — so the inference itself has to stay inside it. Most small businesses don't need this yet. When the work calls for it, we build it.
Where it pays off
Ten jobs an agent does well — and the form that fits each.
From the first hello to the renewal — across support, sales, account management, and how knowledge moves through the team. Which form fits comes down to the data the job touches.
Customer support triage
Glass · two roomsAnswers the common questions from your SOPs and routes the rest to a person, context already gathered.
Outreach that explains itself
Glass · two roomsA prospect sees the exact record you hold, asks why you reached out, and sets their own status.
Lead qualification
Cloud agentGreets an inbound lead, asks your playbook's questions, then routes or books — not a generic script.
Quote & RFP first drafts
Your cloudAssembles a first pass from your product, pricing, and approved language — a human signs, nothing auto-sends.
Account-manager call prep
Your cloudA one-screen briefing — account history, open items, the right play — before the rep walks in.
Live battle cards
Cloud agentServes the current competitive card on demand. Update it once; every rep has it the same hour.
Objection handling on tap
Cloud agentA rep asks “how do I answer X?” and gets the approved, sourced talk track — the same answer team-wide.
New-rep onboarding & training
Cloud agentGrounded in the playbook and product, so a new hire can ask anything and role-play objections.
Customer onboarding & activation
Glass · two roomsWalks a new customer through setup from your runbook, tracking what's done against their own record.
Knowledge that compounds
Shared memoryWins, losses, and key calls captured back into one shared memory — current for everyone, not trapped in inboxes.
Notice the pattern: it's one shape, pointed at different jobs. Swap what the agent is handed and what it remembers, and the same architecture serves the whole funnel.
How it remembers
An agent is only as good as what it knows.
A memory layer runs through every agent above — it's the durable knowledge the agent works from. Which kind you want depends on how your knowledge is shaped — and it's often a blend, with more than one of these working together in the same system.
- 1
A curated wiki
Durable knowledge written in plain markdown — vetted, human-authored, and governable. The flavor we prefer (it follows Google's Open Knowledge Format); this site's own knowledge runs on one. Every answer traces back to something a person approved.
Sounds like you if: your knowledge is already written down, and you want answers you can trust and source.
- 2
A vector database
Similarity search over chunks of text. You hand it a large pile of documents and it finds the passages closest to a question, even when the wording doesn't match.
Sounds like you if: you have a big, messy pile of documents and need fuzzy search across all of it at once.
- 3
A knowledge graph
Entities and the relationships between them, mapped explicitly. The structure carries meaning the raw text can't — who connects to what, and how.
Sounds like you if: the connections between things matter as much as the things themselves.
The one we're known for
Context, not access.
The governed two-room form is the one people come to us for — we call it a glass agent. It holds no keys, no database connection, and no way to reach into your systems. Before it answers, a trusted broker assembles a small packet — the caller's own record, plus the right slice of the memory layer for the job — and hands it in through code. The agent works only from what it was given, never reaching for more.It can't leak what it was never handed.
That's the honest version of what the big managed platforms do under the hood, built inside your own boundary, where the proof lives with you. So “the agent can't reach our data” stops being a promise on a trust page and becomes the shape of the system.
Keeping it healthy
An agent isn't finished at launch.
Two things move under every agent. Your business changes — a process shifts, a document goes stale — and the model itself gets better, which can make yesterday's guardrails the wrong ones. So the agent and the harness around it both need tending.
That's the part we stay for. We watch what the agent reads, what it's allowed to touch, and whether its answers still hold up — and we adjust before drift turns into cost. A launched agent that nobody tends doesn't fail loudly; it just quietly stops being right.
Recognize your situation?
That's the conversation. Tell us the job and the data it touches, and we'll tell you the simplest form that does it safely — no pitch, just a straight answer.