Design
Model the work before the agents do it.
The most expensive bug in an agentic system is a bug in the workflow design. Models can be swapped. Prompts can be rewritten. But if the agent is solving the wrong problem — or solving the right problem in a shape that doesn't fit the people using it — no amount of prompt engineering will recover the engagement.
Design is where we earn the trust to build. We don't open an editor until the system is mostly designed on paper.
What you get
Workflow models
Diagrams of the work as it is today, the work as it would be with agents, and the human-in-the-loop touchpoints between them. Includes escalation paths and recovery flows.
Tool / MCP catalog
A versioned catalog of every capability the agent can invoke, with input/output contracts, side-effect documentation, and access controls. The catalog is the agent's surface area. We treat it like an API.
Prompts as artifacts
Prompts in a repository, versioned, reviewed, and tested. Each prompt has a purpose statement, expected inputs, expected outputs, and a small suite of behavioral evals.
Eval design
What "good" means for this agent, expressed as a set of tests that run before deployment and continuously after. Without this, you have no way to know your system is working. With it, you have a release process.
Operator UX
The interfaces your team will use to monitor, override, and improve the agent. Often more important than the end-user UX.
Engagement shape
Four to ten weeks, often concurrent with Build.
Senior designer lead, embedded prompt and evals engineer, design partner from your team.
What this is not
This is not a UI agency that puts a chat box on top of an LLM. This is the design of the underlying system, the contract between the agent and the world.