— Practice / AI agents
AI agent development.
Anyone can wire a model to a prompt. An agent that does real work — calls tools, retrieves facts, stays honest under load — is an engineering problem. We build production agent systems: the retrieval, the evals, the guardrails, and the boring infrastructure that keeps an agent from looping or lying.
— Why us
Agents, not demos.
We run a grounded, cited AI search on this site, publish open-source templates for RAG and MCP that other teams deploy, and expose our own tools to agents over a live MCP server. We build for clients the same way — engineered, measured, and observable, not bolted on.
— What we build
The parts that come after the prompt.
Tool-using agents
Agents that call your systems to get work done — Anthropic and OpenAI SDKs with function calling, and the orchestration that keeps a multi-step task on the rails.
Production RAG
Retrieval that grounds answers in your own content with citations — chunking, embeddings, and a vector store tuned to the actual corpus, not a tutorial.
MCP servers
Give your agent — or someone else’s — a safe, standard way to reach your software as tools. We build the server and the agent that uses it.
Evals as a habit
If you cannot measure regressions, you have a prompt, not a system. We write evals into the pipeline so a change that breaks the agent fails the build.
Guardrails and observability
Input validation, output checks, rate limits, and structured logging — so you can see what the agent did and trust what it does next.
The right surface
API, embedded UI, an MCP server, or a background worker — we build the agent into the surface that fits the job, not a chat box because it is easy.
— How it works
The engagement.
Same five-step method as every SetKernel build — Brief, Architect, Sprint, Ship, Operate — each with a written artefact you review. We start from a short written brief: the task the agent should own, the systems it needs, and what a good answer looks like. You get a scoped price and a fit / no-fit answer within one business day.
— Where we work
Atlantic Canada, and worldwide.
We are a complete technology partner in Halifax, Nova Scotia, Canada, and we work remotely with teams well beyond the region. Agent systems are cloud-native by nature — where your team sits does not change the build.
— Questions
Before you write.
What counts as an "AI agent"?
Software that uses a model to decide and act — it calls tools, retrieves information, and takes multiple steps toward a goal, rather than returning a single completion. The engineering is in the tools, the retrieval, and the checks that keep it reliable.
Do you train or fine-tune models?
No. If your problem genuinely needs custom weights we will point you to people who do that for a living. Most production value comes from retrieval, tool use, evals, and orchestration around strong existing models — which is what we build.
How do you keep an agent from going wrong in production?
Evals in the deploy pipeline, input and output validation, rate limits, and structured logging you can audit. We treat an agent like any other production system — measured and observable — not a black box you hope behaves.
Can we see your work first?
Yes. Use the grounded, cited AI search on this site, read our open-source RAG and MCP templates and their companion essays, and install our live MCP server. We would rather show you working systems than a deck.
How do we start?
Send a short written brief — the task the agent should own, the systems it needs, the deadline. We reply in writing within one business day with fit / no-fit and, if fit, a scope and price. No discovery call before the brief.
— Engage
Have a task you want an agent to own?
Tell us in two paragraphs — the task, the systems, what a good answer looks like. We reply in writing within one business day.