AgenticOS

AgenticOS is where enterprise teams build AI agent workflows: a support agent that reads a ticket, checks a policy and drafts a refund, with a manager’s approval in the middle. The first version worked, but it looked and felt like a set of admin tables and dense modals. The most important safety feature, Dry Run, was a small switch inside a long form.

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, because of rising costs, unclear value and weak risk controls. I treated those three reasons as the design brief. If the product could make cost visible, value provable and risk controllable, it would get approved. If not, it would end up in the 40%.

I led the UX vision and process end to end, with two product designers, a UX researcher and a content designer. One designer owned Build (canvas, agents, templates). The other owned Run, Observe and Govern. I owned research strategy, the information architecture, the design system direction and the handoff, and I kept the decision log: what we chose, why, who decided, and what we did not choose.

The design system started in week four, not week eight, so screens were built on it instead of before it. A steering group of the CPO, CTO and VP Sales signed each gate, and a one-page decision memo went out the day before every review. Short memos made for short meetings.

I began with a screenshot audit of twenty screens from the existing dark-mode product, rated P0 to P2. The base was solid, but the product’s promise was invisible. Dry Run was a tiny switch in a dense form. Mission Control showed four failed KPIs while every service tile said Healthy. The task builder was a list, not a canvas, and nothing taught people the five-level model of Department, Task, Task Pipe, Goal and Run.

Then I mapped how a team builds one AI workflow today. It touches eight places: Slack and Jira for the idea, scripts or n8n to build, an LLM console for the AI logic, a laptop with copied data to test, email and meetings to approve, CI to deploy, log tools to monitor, and spreadsheets to report.

Six proto-personas covered the whole life of a workflow: Priya the business analyst who has the idea, Sarah the process architect and Karthik the AI engineer who build it, Rahul on call who keeps it running, Meera in risk and compliance who approves it, and Alex who pays for it.

We scored all thirteen bets on user impact, business value, effort and risk. Release one had a simple rule: one complete, safe journey. Canvas, Copilot drafts, Sandbox, Replay, Approval Inbox, environments, Health and Recovery, and basic policies. The full Policy Studio, Quality Gate, Cost and Value and Blueprints came next. Co-editing and the marketplace came later.

The brand idea was calm control. Quiet surfaces instead of loud gradients, plain words instead of jargon, and proof instead of "just trust the AI". Agent Indigo is used for action and focus only, roughly 5–8% of any screen, so that when something is indigo it actually means something.

Thirty-four screens in eight groups, designed at 1440 and responsive down to 1280: Foundation, Build, Run, Observe, Govern, Connect, Admin and Global. A few of them carry most of the story.

Screens were marked "Ready for dev" in Dev Mode, with tokens as variables, the variant library, a canvas interaction spec (drag, connect, zoom, keyboard), a copy deck, annotated flows and accessibility notes per component. We walked engineering through each area, answered Jira questions within a day, and ran design QA on staging every sprint.

"Done" had a strict meaning: every state designed, every string final, every token used, and every screen checked against the six principles.

If I did it again, I would get two design partners into the sandbox flow even earlier, with real data, before the canvas was polished. The riskiest assumption in the project was that risk leads would approve inside the product instead of over email. That deserves real evidence first.