An Open-Source Agent Stack for Product Teams
A practical stack using browser agents, memory, skills, evaluation, and support bots for teams building real agent workflows.
A product team does not need the biggest agent framework first. It needs a stack that turns one useful workflow into something measurable and maintainable.
A practical open-source stack can start with Webwright for browser tasks, AgentMemory for durable context, Garden Skills for reusable workflows, MLflow for evaluation, and Tiledesk when the agent touches customer support.
Layer by layer
- Agent runtime: choose the smallest framework that can complete the task.
- Tool boundary: expose only the actions the agent needs.
- Memory: store durable project context, not secrets.
- Skills: package repeatable expert workflows.
- Evaluation: measure the same task every time.
- Handoff: define where a human takes over.
First milestone
Pick a workflow that matters but will not cause damage if it fails. Build a repeatable test set before adding more autonomy.
What to avoid
Avoid stacking five new systems before you can measure one successful workflow. Agent stacks become useful when each layer has a job.