Atlas.
A control room for AI agents — when one agent becomes a fleet.
- What
- A control room for AI agent fleets: morning brief, traces, evals and billing.
- Why
- Once a team runs ten agents, visibility collapses: logs scattered, evals too late.
- My role
- Product strategy, UX and visual design, 0→1 to a Figma prototype.
- Result
- 30 screens (20 desktop · 10 mobile) · 10 s to read the morning brief
Morning briefs, traces, evals and billing for agent fleets — in an editorial interface that makes complexity feel manageable.
Read the full story
As AI agent fleets grow from one to dozens, teams lose visibility. Atlas is a full-stack observability and operations platform — morning briefs, trace explorer, eval pipelines, billing oversight, and multi-tenant collaboration — wrapped in an editorial interface that makes complexity feel manageable.
Warm paper. Serif data. Rust alerts.
One agent takes an afternoon. Ten take a control room.
No fleet overview
Status, spend, evals — scattered.
Traces buried in logs
Raw JSON, no timeline.
Evals happen too late
Regressions reach production.
Read the full story
Teams deploying AI agents have no control room. You can deploy an agent in an afternoon. But once you have ten — across different workflows, prompts, and cost centers — visibility collapses. Logs are scattered, evals live in spreadsheets, billing is a surprise at month end.
No unified fleet overview. Teams manage agents across multiple tools. There’s no single place to see all agents, their status, spend, and eval health at once.
Traces buried in logs. Debugging a failed run means trawling raw JSON logs. There’s no structured trace view — tool calls, LLM outputs, branching decisions — presented as a readable timeline.
Evals happen too late. Quality checks run manually, or not at all. By the time a regression is caught, it’s already in production. A continuous eval pipeline with pass/fail trends was missing.
The morning brief. Fleet at a glance.
Why it opens here
Every team lead opens Atlas to the same view: agent fleet health, today’s spend vs. budget, run volume, and eval pass rate — all above the fold. The card-based layout gives each agent a name, status badge, and burn rate.
Agent naming (Wren, Tarn, Lev, Sole…) was intentional: names are easier to remember and talk about than UUIDs in incident response.
Deep dive on any single agent.
Read more
Click into an agent and get the full picture: live status, model version, system prompt, run history, cost breakdown, and recent eval scores — all in one scrollable view.
The prompt panel is editable inline: teams can draft and push prompt changes without leaving Atlas, and each version is logged for diffing.
Every run, fully observable.
Read more
The trace view renders a complete execution timeline: each tool call, LLM inference, retrieved context chunk, and branch decision — laid out chronologically with latency and token counts.
Clicking any step shows the raw input/output payload. Engineers can pinpoint exactly where a run diverged from expected behavior without leaving the UI. Execution timeline · tool call inspector · token counts · latency breakdown.
Catch regressions before prod.
- 01Auto-run on deploy
- 02Pass-rate trends
- 03Threshold alerts
Read more
Every agent has an attached eval suite. On each deploy or prompt change, Atlas runs the suite automatically and plots pass/fail trend over time. Threshold alerts fire if pass rate drops below the team’s SLA.
The eval grid shows individual test cases with expected vs. actual output — designed for non-engineers to review quality without reading raw logs. Test case explorer included.
Run management across the whole fleet.
Read more
The Tasks view is the ops center: filter runs by agent, status, date range, or cost. Bulk-cancel stuck runs. Inspect any task inline. Export a filtered dataset for analysis.
The task detail slide-over shows full metadata without leaving the list — a pattern that keeps the overall spatial model stable when drilling into specifics.
20 desktop screens. Onboarding to billing.
The full list
Complete product coverage: onboarding, fleet management, agent config, trace, tasks, evals, billing, audit log, knowledge base, triggers, prompts, notifications, settings, and collaboration.
Home — morning brief · Agent list · Agent detail · Trace explorer · Tasks · Task detail · Evaluations · Billing · Audit log · Knowledge base · Triggers · Prompt editor · Notifications · Settings · Share — collaboration · Sign in · Empty state · First agent live · Rate limit warning · Timeout error.
Even errors read like a page.
10 phone screens. Monitor and approve anywhere.
Read more
The mobile experience focuses on monitoring and approvals — check fleet status, review task alerts, and approve or cancel runs from anywhere. iOS-native feel at 390px.
Why editorial, not dashboard-grey.
“Wren is stuck” — not “agent-7f3a2b is stuck.”
Fraunces for data, not just headlines
Morning brief as the entry point
The three decisions
Most ops tools default to cold blue-grey SaaS chrome. Atlas chose the opposite — warm paper tones, serif typography, rust accents — to make intensive monitoring sessions feel less fatiguing.
01 · Fraunces for data, not just headlines. A variable optical-size serif reads equally well at 124px display headings and 13px table cells. The editorial feel signals “this tool has opinions” — it’s not a generic dashboard.
02 · Named agents, not UUIDs. When something goes wrong at 2am, you want to say “Wren is stuck” not “agent-7f3a2b is stuck.” Short, memorable names reduce cognitive load in incident response and daily standups.
03 · Morning brief as the entry point. The home screen is a daily digest, not a live feed. Key metrics are readable in 10 seconds. Full detail lives one click away — not immediately visible, avoiding dashboard anxiety.
What was designed.
Desktop screens
Figma, production-ready.
Mobile screens
iOS-native, 390px.
Design language
Fraunces + Inter Tight.
Complete product
Strategy + UX + visual.
Read more
20 desktop screens — full product coverage from onboarding to billing; Figma, production-ready. 10 mobile screens — monitoring and approval flows on the go; iOS-native feel, 390px. 1 cohesive design language — warm editorial, Fraunces + Inter Tight; tokens, dark mode, responsive. 0→1 — a complete product designed from concept to full prototype, no prior art; strategy + UX + visual.
