Built for a metered API budget, not unlimited compute. Every agent works against its own daily cap, enforced automatically.
Agent statuslive
Task queue0 pending
Agent activity
Ops logrecent
All agents8 agents · shared SQLite task queue
Task queueDispatcher-assigned, newest first
AnalyticsSynthetic session data
Agent activitylast ~7 min
UpcomingIdleReviewingWorking
Model routing split
0
tasks on Sonnet
0
tasks on Haiku
Sonnet: client-facingHaiku: low priority
Completions
0
last interval
0
total this session
last 12 intervals · 5s each
Within budget
8
under 80%
Near limit
0
80–99%
On hold
0
at 100% cap
Ops logHealth, budget and dispatch events
Safety railsAlways on, never disabled
Autonomy isn't the same as no oversight. These constraints run underneath every agent, all the time. They're not configurable per task, and no agent can reason its way around them.
Irreversible actions require a human "yes"
Any deletion, overwrite, database drop, or migration stops completely and waits for explicit approval. No agent proceeds on its own judgment.
Fix attempts are capped, not infinite
The same bug gets a maximum of two automated attempts before the agent stops and escalates with a diagnosis, instead of retrying forever.
Production requires four-way sign-off
QA, design review, architect review, and an explicit human "go": all four, no exceptions, before anything ships.
Spend is append-only and never silently expanded
Every request is logged in an append-only ledger (no edits, no deletes) and follows a staged trust model rather than blanket autonomy.
High-risk automations default to dry-run
Anything that can act in the real world (an external message, a production deploy) starts simulated and needs explicit written approval to go live.
Agents don't touch what's outside their task
File and system access is scoped to the current task. Nothing gets modified outside that boundary, and it's enforced across the whole team, not self-policed.
Sample outputsA taste of what agents actually produce
Everything in here is an illustrative example written for this demo, not a real commit, message or report. It's here to show the kind of output an agent hands back, not to be exhaustive. Use the arrows or the dots to move between agents.
KA
Kayla
Chief Architect · illustrative example
Coordinates the roster, prioritizes the queue, and reports status back.
Output: status digest
6
tasks completed overnight
2
flagged for review
0
blocked
Morning digest: 6 tasks completed overnight, all agents green. Baljit's booking-card fix is ready for review. Fabien flagged a slow query on the export endpoint, added to today's queue as high priority. Anika's daily scan is in too, one item worth a look on the pricing page.
Good, let Fabien's fix go out today, no need to wait for review. Take a look at Anika's pricing note yourself and tell me if it's worth acting on.
Done. Fabien's fix is merged and queued for deploy. On pricing: agreed, it's a real signal, not noise. I've added "review pricing page copy" to Lottie's queue as medium priority.
AN
Anika
Research & Intelligence · illustrative example
Runs research and market scans, and flags what's worth acting on.
Output: research note · 4 signals flagged
Competitor pricing shifted roughly 8% lower across two comparable tools this week, worth a pricing page review.high
Recurring theme in support threads: users want CSV export, not just PDF.med
A new entrant launched a similar tool this week, priced about 20% below us, worth a feature comparison, not necessarily a price response.med
Search interest for booking-software terms is up seasonally, worth timing any pricing change around it, not against it.med
Read together: the pricing pressure is real, but the new entrant undercutting us is unproven on feature parity. I'd hold price and lead with the CSV export request instead. It's the complaint that's actually recurring.
fix(api): add query timeout and an explicit empty-state response for slot lookups
GE
Georgia
Sales & Marketing · illustrative example
Researches prospects and drafts outreach.
Output: outreach draft
Subject: quick question about your booking workflow. Hi [Name], noticed your team's still coordinating slots by email. We built something that handles that end-to-end, happy to walk through it in 10 minutes if useful, no pressure either way.
status: drafted, queued for review before send
PR
Priya
QA & Testing · illustrative example
Tests everything before it ships.
Output: test report
✓Booking flow: 42 / 42 passing
✓Auth edge cases: 18 / 18 passing
✓Payment webhook handling: 11 / 11 passing
✗Export endpoint: 1 failing (timeout under load), flagged to Fabien
stack trace
TimeoutError: request exceeded 4000ms
at getAvailableSlots (api/bookings.ts:142)
at Object.test (export.spec.ts:58)
LO
Lottie
UX/UI Designer · illustrative example
Proposes and refines the interface design.
Output: design proposal
Proposed simplified layout for the settings panel: fewer nested sections, primary actions surfaced at the top.
Before
→
After
Why
Collapse five nested tabs into two top-level groups
Move Save and Cancel out of a scrolled footer and into the header
Surface the one setting most support tickets ask about, instead of burying it
JO
Jorge
Systems & 24/7 Ops · illustrative example
Watches the whole system's health, 24/7.
Output: health report
99.7%
uptime, 30d
1
auto-restart, last 7d
03:14✓All 8 agent processes healthy, heartbeats on schedule
03:41⚠Georgia's session missed one heartbeat, restarted automatically, resumed in 4s
This is a simulation. Agents, tasks and figures are synthetic and regenerate on load. No client data, real spend or production system is connected. It's styled to match Everant Studio's real internal console and mirrors its actual mechanics: model-tiered routing, per-agent usage gates with an 80% alert and a 100% hold, and automatic recovery when an agent misses a heartbeat.
This console is a live simulation of Everant Studio's real internal dashboard: an 8-agent studio I run myself, coordinated through a shared task queue instead of a group chat. It's a light version: it mirrors the real mechanics (budget gates, model routing, heartbeat recovery), but runs entirely on synthetic data, and a lot of the real system (client work, spend, integrations) has been left out.