01
Map the workflow
Start with the user task, product context, and decision the agent is meant to support.
Design the system around the AI agent.
A design engagement for product teams that have started building an AI agent but need clarity on the workflow, boundaries, evaluation, and oversight before adding more implementation.
For agents that must work reliably inside an existing product or system.
Discuss an AI agentA model is not the whole system.
The prototype can produce an answer, but nobody can state which actions it is allowed to take.Relevant product knowledge is scattered, stale, or exposed to the agent without clear boundaries.The team cannot say how a correct result will be evaluated or what should happen when confidence is low.Retries, failures, escalation, and human review are being decided during implementation rather than in the design.
What the engagement covers.
The work makes the surrounding system explicit before implementation decisions become expensive to undo.
02
Bound the system
Define knowledge access, tools, permissions, failure behavior, and human intervention.
03
Design evaluation
Specify what good looks like, how to detect failure, and what evidence is needed before release.
What you leave with.
A bounded workflow
The agent's inputs, decisions, actions, and stopping conditions made explicit.
Tool and knowledge boundaries
A design for retrieval, permissions, tool access, and the information the agent may use.
An evaluation and oversight plan
Clear checks for output quality, failure paths, and when the workflow returns control to a person.
An implementation path
A sequenced build plan with assumptions, open questions, and the work required to move forward.
A public example of the approach.
Pratrol is an open-source pull-request triage system. Its design makes routing decisions and their rationale inspectable rather than treating automation as a black box.
Public work
PR Triage System
Read the system design or inspect the source code.
Need implementation after the design?
The engagement produces the system design. Hands-on implementation and hardening are scoped through delivery.