Agents that ship, and stay in your control
This is kept separate from the app work, with its own standard, because the failure modes are different. We build production agent systems: the orchestration, the tools, the retries, and the guardrails that decide what an agent is allowed to do on its own. A human approves every irreversible step.
An agent that works, and asks before anything risky
It reads and analyzes on its own. Every irreversible action, moving money or emailing a customer, stops and waits for you.
Try denying one. The agent skips it, flags it for a person, and carries on with the rest. That is the whole idea, and it is easier to see than to describe.
A faithful walkthrough of the approval flow, running in your browser. It calls no model and sends nothing anywhere. The real gate is Gatekeeper, which does this against live tools.
Orchestration
Multi-step work that survives contact with reality: a plan, tools that actually run, and a runtime that knows what to do when step four fails at 2am.
- Planner and worker separation, so a bad plan does not become a bad transaction
- Durable state, so a crash resumes instead of restarting
- Observability: every decision the system made, and why, readable after the fact
Guardrails
The part most demos skip. We define, up front, the set of actions an agent may never take alone, and build the gate that holds it.
- A human approves anything irreversible: money moved, messages sent, records deleted
- Budget ceilings enforced in code, not in a prompt
- Refusal to act on low confidence, with the ambiguity handed back to a person
Private by default
A specialty in agents that keep company data inside the company. Where a workload can run on hardware you own, we will tell you, and build it that way.
- On-device and self-hosted models where the task allows it
- Clear boundaries for what, if anything, reaches a third-party model
- No training on your data, ever, by us or by anyone downstream
Evaluation
An agent that works in a demo and fails in production is the normal outcome. We measure before we ship and keep measuring after.
- Task suites built from your real cases, not synthetic ones
- Cost and quality tracked per run, so a model swap is a decision with numbers
- Regression gates: a change that makes the system worse does not deploy
Our own agent products
We run this stack on our own work first. LuciLead is the current in-house build.
LuciLead
Outbound that writes itself: find the right prospects, enrich them from independent sources, and draft outreach that sounds like a person wrote it. Every claim in a draft has to be corroborated by at least two sources before it ships, which is the same rule we hold client agents to.
Human in the loop
No system we build takes an irreversible action without a person approving it. That is a design constraint, not a setting you can turn off later.
The practice, not the pitch
Agent shops are easy to claim and hard to verify. These three are the parts of our own stack you can drive yourself, right now, without an account.
What a build actually includes
Agent projects go wrong in the same few places every time. These are the parts that get built whether or not they were asked for.
The runtime
- Planner and worker separation
- Durable state across crashes
- Retries with real backoff
- Tool definitions you can read
- A full trace of every decision
The gate
- An approval step for anything irreversible
- Spend ceilings enforced in code
- Refusal on low confidence
- An audit log that survives a restart
- A kill switch that actually stops it
The proof
- Task suites from your real cases
- Cost and quality tracked per run
- Regression gates before deploy
- A baseline you can argue with
- Handover docs, not tribal knowledge
What people ask about agents
Most of these come up before the word “architecture” is ever said, and they matter more.
Does my data go to a model vendor?
Only where you decide it should, and the boundary gets written down before anything is built. Where a workload can run on hardware you own, we will tell you and build it that way. Nothing you give us is used for training by us or anyone downstream.
What stops it doing something expensive?
Budget ceilings live in the code path, not in a prompt asking the model to be careful. A prompt is a suggestion. A hard limit in the runtime is not.
How do I know it works before it is loose on real data?
An evaluation suite built from your actual cases, run before and after every change, with a gate that refuses to deploy a regression. “It looked good in the demo” is not a measurement.
Should this even be an agent?
Often no. A scheduled script, a form, or ten lines of SQL beats an agent for a surprising share of the workflows people bring in. You will get that answer if it is the right one.
Tell us the workflow you wish ran itself
Bring the process that eats your week. We will tell you honestly whether an agent should touch it, and what it would cost to do properly.
Describe the workflow