Cloud platform · Agentic AI operations
The problem: agents ran from pasted prompts, automations fired unwatched, knowledge went stale in silence and token spend showed up as a single invoice line. No owner, no state, no clock.
The work: one multi-tenant cloud console where every agent, workflow and knowledge source has a single trustworthy state, a named owner, a success rate and a cost per run — scannable as rows, reasoned about on a canvas, and reachable in two keystrokes.
Teams had already adopted AI assistants and automations. What they did not have was a place where those things were owned, measured or safe to change.
Prompts were pasted into notebooks and chat threads. When an assistant started answering badly, no one knew who owned it, which model it ran, or when it last changed.
Workflows fired on webhooks and schedules with no shared view of runs, success rate or duration — the first signal of failure was an angry customer.
PDFs, docs and help-centre pages were indexed once and never revisited, so agents quoted retired pricing with total confidence.
Token spend arrived as one invoice line. Nobody could attribute it to an agent, a workflow or a bad prompt.
When an agent handed a conversation to a human, the trail ended. Escalated threads aged with no owner and no clock.
Four roles share the same console with very different jobs — building, operating, curating and accounting for the workforce.
“Which of my 40 workflows is failing right now?”
“I want to change instructions without breaking production.”
“Is the answer the agent gave still true?”
“What is this workforce costing, and who is accountable?”
Fourteen sessions with operators, builders and knowledge owners — plus a week reading run logs and support threads.
Every agent, workflow and source resolves to exactly one badge — Active, Paused, Draft, Error, Indexed, Processing, Failed, Syncing. Ambiguous states were the top source of mistrust in research.
The same feature needs two surfaces: a dense table for scanning hundreds of runs, and a node canvas for reasoning about one flow's branches.
Showing $/run on the same row as success rate changed decisions instantly — teams paused expensive, low-value workflows within minutes of first use.
Power users asked for one command surface over deeper menus, so ⌘K covers recents, create actions and navigation in a single modal.
One sidebar, two groups: Workspace for the work, Settings for the guardrails. Every destination answers a single question.
The morning read: active agents, conversations today, success rate versus last week and AI cost, over one shared 12-month axis.
A catalogue of the workforce — status, model, sources, tools, conversations and success rate, with Test and Configure on every card.
Workflows as both a live canvas preview and a dense run table filtered by All, Active, Paused and Draft.
Sources and collections with index state, chunk counts and last-indexed time, plus drag-and-drop ingestion.
Threads, escalations and performance reporting for the whole workforce.
⌘K over recents, create actions and navigation, so nothing needs more than two keystrokes.
Every surface below is shown as it ships, with the problem it was built against and the design move that answers it.
Teams started the day in five tabs and still could not say whether the workforce was healthy.
Four tiles answer volume, quality and cost first; the activity chart, agent list, human activity feed and automation runs sit beneath on one screen — every number carries its own change-versus-last-period.
The first build read as marketing-calm; operators on wall displays lost the signal from across the room.
A second density and contrast pass: tighter sidebar rhythm, stronger accent on the live series and full-width tiles — same information architecture, readable at three metres.
Workflow failures, indexing results and usage limits arrived by email, hours late, to whoever happened to be on the thread.
One panel groups today's events, marks unread with a dot, and mixes system signals (usage at 80%, maintenance) with work signals (workflow completed, knowledge indexed) so nothing needs a separate inbox.
Nobody could list the agents in production, let alone their model or owner.
Cards carry status, model, source and tool counts, plus three numbers that matter — conversations, success rate, last active — with Test and Configure inline so a fix never needs a detour.
Prompts were edited blind: no version, no environment, no way to tell whether a change helped.
Performance over 30 days sits above the instruction editor, and the right rail pins the live switch, version, environment and model settings — so the person changing behaviour sees the consequences in the same viewport.
A workflow was either a diagram nobody could audit or a log nobody could picture.
The featured flow renders as a canvas with run stats beneath it (last run, success rate, average duration, cost per run); everything else lists as rows that stay put while numbers update.
Conditional logic lived in code and in one engineer's head.
Typed, colour-coded nodes and an explicit condition branch make the logic legible; Draft, Autosaved, Test and Published sit in one header row so publishing is a deliberate act, never an accident.
Half-built automations sat in the same list as production ones and got triggered by mistake.
Draft is a first-class filter: unpublished work keeps its metrics and stays visible, but is clearly separated from anything that can fire.
Indexing was a black box; stale and failed sources looked identical to healthy ones.
Every source shows Indexed, Processing, Failed or Syncing with chunk count, collection and last-indexed time — so the answer to "is this still true?" is on the row, not in a log.
Creating an agent or workflow meant three clicks through a growing sidebar.
⌘K groups Recent, Create and Navigate, with arrow-key select and enter-to-open — the fastest path for the operators who live here all day.
I led UX end to end: research, IA, interaction design, the state and density system, and the spec engineering built from.
Agents, workflows and sources share a single state language and colour mapping. Learn it once on the dashboard and it holds everywhere.
Tables never re-order themselves as values change, and the canvas is reserved for structure — the two surfaces never compete.
$/run and daily AI cost sit beside success rate rather than in billing, so quality and spend are traded off in the same glance.
Autosave protects the draft; Test and Published are separate, deliberate actions. Nothing reaches production by autosave.
Every list, chart and export renders after the caller's workspace and role are resolved, so one console safely serves many organisations.
An escalated conversation carries an owner and a clock, so agent-to-human handover has the same accountability as the rest of the queue.
Measured across the first two releases with the pilot operations team.
What I would carry into the next version of this product.
Every research round asked for more on screen, not less. The second, higher-contrast dashboard was the version operators kept.
Grouping by object and severity, plus an explicit acknowledge loop, is the obvious follow-on once the workforce passes fifty agents.
Putting spend next to success rate produced decisions on day one — a reminder that the smallest number on a row can be the most persuasive.