CEDX Agents runs crews of routers, planners and specialists against your tools. The canvas decides what an agent may touch — reads run automatic, writes wait for a human — and every run lands in a fleet table with a success rate, a score and a cost.
The demo crew on the Studio canvas is estimated at $0.0290 a run, 2.1 seconds, 8 steps — before it runs. Two guardrails ship in front of it: deny unlisted tools, redact card numbers. Autonomy here is a configuration, not a mood.
agents.cedxsystems.com — live build
Runs on demo data. The budget card says “illustrative from the demo sample” on the screen itself.
72 agents in the fleet323,744 historical runs behind their scores
48 tools, each one watchedfail rate, P50/P95 latency and cost per call
Writes wait for a humancrm.update on the demo canvas is marked require-human
What it is
Three things stand between an agent and an incident.
The crew is a diagram you can argue with
The Studio canvas for the demo crew has 7 nodes and 8 edges: triage.router on assist-balanced, refund.drafter on assist-fast, a policy branch on “intent = refund?”, and two guardrails in front of everything — guard.allowlist denies unlisted tools, policy.pii redacts card numbers. The run is priced before it happens: $0.0290, 2.1 seconds, 8 steps.
7 nodes, 8 edges, 2 agents, 2 guardrails — all on screen
Read tools run auto; crm.update is write · require-human
Run preview streams a seeded trajectory
agents — screen-2
A fleet table, not a faith exercise
72 agents with version, role, crew, model, runs, success, score and cost per run on every row. triage.router is on 2.4.0 with 18,200 runs at 94.0% success; onboard.detect.exp scores 97 at $0.3745 a run. Cost and quality argue in public, from the same table.
Roles are explicit: Router, Planner, Detector, Specialist, Remediator
Success and score are separate, sortable columns
Tag chips — account, capacity, faq, latency, policy
agents — screen-3
Tools are the failure surface
48 tools, 685,492 historical calls, and a fail rate on every one. The view sorts by failure: vector.upset.core is critical at 10.7%; crm.update is degraded at 8.40% over 42,100 calls. Twenty tools sit at degraded or worse, and the header counts them as needing attention.
healthy · degraded · critical are statuses, not adjectives
P50 and P95 latency next to the fail rate
Cost per call down to $0.0002
agents — screen-4
Product tour
Four screens, captured from the running build.
Not a mockup and not a concept deck. This is what opens at /app/agents.
agents.cedxsystems.com
01 — Overview
Agent ops as a ledger, not a pet project
320 runs in the sample — 265 ok, 34 error, 12 timeout, 9 cancelled — with $1.2K of sample spend against a $48K month budget paced at 38%. The alerts arrive root-caused and priced: crm.update's failure spike is traced to an unknown field 'prio' in the signature, at $420.
Tool fail rate 8.5% is a headline card, not a footnote
Pareto cells: 5 of 54 scored
Budget pace 38% · “illustrative from the demo sample”
02 — Studio
Design the crew, price the run
Agents, branches, tools and guardrails are nodes you drag. The demo crew estimates at $0.0290 a run with a 2.1-second, 8-step trajectory, and the preview streams that seeded trajectory before anything touches production.
Est. cost/run and trajectory on the header
Guardrail nodes sit in the graph, not in a policy PDF
+ Add tool node — the canvas is the build surface
03 — Fleet
Seventy-two agents, all of them graded
Every agent row carries its version, role, crew and model with runs, success, score and cost per run. support.escalate holds 97.0% success at $0.0020 a run; ops.detect.v1 scores 92 at $0.2131. Both facts are visible, which is what makes either useful.
323,744 historical runs behind the scores
Average success 85.6%, average score 78 across the view
CSV of the filtered view
04 — Tools
Watch the hands, not the face
Agents fail through their tools, so the tools get their own screen: calls, fail rate, P50 and P95 latency, cost per call. file.read is critical at 7.80% over 27,488 calls — that is where an agent incident actually starts.
685,492 historical calls in the filtered set
Average fail rate 3.71% across the view
20 tools at degraded or critical, counted up top
Who runs it
Three roles keep a crew accountable.
Roles, not references. We have no named customers yet, so nobody in these photographs is quoted, credited or claimed as one.
Agent operations
Owns the fleet table: watches success and score by crew, and reads the 34 error runs in the sample before anyone writes a postmortem.
72 agents · 85.6% avg success
Platform engineering
Owns the tool surface — 48 connectors with fail rates and P95s — and fixes the schema when crm.update starts rejecting an unknown field.
48 tools · 20 need attention
Support operations
The human in require-human. Approves the writes the refund.drafter queues, and keeps the tool allowlist honest.
writes · require-human
The shape of it
What the demo workspace actually looks like.
Every figure below is legible in the captures above. Nothing here is a projection of your estate — it is the state of the demo data.
72agents in the fleet323,744 historical runs
48tools agents can call20 degraded or critical
265ok runs of 320 in view34 error · 12 timeout · 9 cancelled
$0.0290estimated cost per runthe Studio's estimate for the crew on canvas
Root-caused alerts, priced on screensymptom → cause → fix · last 30 days
Pace 38% — “illustrative from the demo sample (320 runs)”, on the screen
How it runs
A crew's life, in the order it actually happens.
01
Design
The crew starts on the Studio canvas: agents, branches, tools and guardrails as nodes, with cost and trajectory estimated before the first run.
02
Preview
Run preview streams a seeded trajectory through the graph — 8 steps on the demo crew — so the plan is reviewed before production sees it.
03
Operate
Runs land in the Overview and the Fleet: 320 in the sample window, with success, score, cost and latency tracked per agent version.
04
Fix
When a tool fails, the alert names the cause — an unknown field, a missing scope, a downstream 502 — and prices it, so the fix is a decision, not a dig.
One record
Crews work the record the estate already keeps.
The demo fleet's crews are named for the work — Support crew, Sales crew, Onboarding crew — because the tools on the canvas are the estate's own verbs.
Finding this out on the third call is worse for you than reading it here, and worse for us.
Agents is not generally available. What opens today is the live build running on demo data — Northline is the software's demo workspace, not a customer.
We have no named customers to show you, so this page shows none.
Autonomy is bounded by design: on the demo canvas, writes sit behind require-human and unlisted tools are denied. We do not claim unattended operation beyond what the guardrails enforce.
The budget figures — $1.2K sample spend against $48K a month — are labelled illustrative on the screen. They are not billing data, yours or ours.
The model lineup in the captures — assist-mini through assist-reasoning and vision-edge — is the demo configuration. Which models a given plan can use is not something we are claiming here.
No audit or compliance certification has been issued for Agents. What we can evidence about hosting, encryption and access is on the security page.
Yes. Every screenshot is a capture of the running build and you can open the same build at /app/agents. It runs on demo data, which the screens themselves say rather than a disclaimer.
What stops an agent doing the wrong thing?
The guardrails on the canvas. The demo crew runs with guard.allowlist — deny unlisted tools — and policy.pii — redact card numbers — and its one write tool, crm.update, is marked require-human. Beyond the canvas, the Tools screen watches fail rate and latency on every connector an agent can call.
What does a run cost?
The Studio estimates the demo crew at $0.0290 a run before it executes. In the fleet, cost per run is printed on every agent — from $0.0020 on support.escalate to $0.3745 on onboard.detect.exp — so the expensive agents are expensive in public.
Can we see failures, or only the wins?
Failures lead. The Overview's headline cards include tool fail rate at 8.5%; the runs-by-status chart shows 34 error, 12 timeout and 9 cancelled runs of 320; and the Tools screen sorts vector.upset.core to the top precisely because it is critical at 10.7% failures.
Is Agents audited or certified?
No certification has been issued. What we can evidence about hosting, encryption, tenant isolation and production access is written up on the security page.
The crew is on the canvas. Run the preview.
Live build, demo data, no card. Then ask whether your writes would have waited for a human.
