Maestro
Live
Marina CostaRede Vida Clínicas

Health

Healthy
0of 100 · +2 this week
Success rate96,4%
Avg. latency2,8 s
Active workflows42
Active agents186
Open incidents3
Workflow success96,4%
−1,2%this week
Active agents186
+12vs. last week
Failed runs248
−18%this week
Compute costR$ 8.421
+6,8%this week

Live agent network

42 workflows · 186 agents · 1 degraded node · updating in real time

Expand

Active incidents

View all
Infinite delegation loopPatient Care · 842 runs affected
Authorization context lostBilling · 128 runs affected
Calendar latency +38%Scheduling · degraded

Cost per agent

Welcome back, Marina

3 open incidents. Want me to explain the loop?

0/300

Workflows

Every patient conversation runs through one of these workflows. Numbers cover the last 24 hours.

Live agent map

Every delegation and context handoff between production agents · in real time

1,0×

Patient Care · Production

size = calls · thickness = handoffs
Live handoffsstreaming · last 60 s
Traces›tr_84921›Handoff

Front Desk → Billing

Context degraded

Handoff inspector · Patient Care · Production · trace tr_84921 · 09:42:18

Handoff path
Front Desksends the context · scope: admin
12,4 KBcontext
420 mslatency
2.431tokens
4 of 5fields
//
Billingrejected the request · scope: patient

Context diff

1 of 5 required fields missing
FieldSent by Front DeskExpected by BillingStatus
paciente_id✓ present✓ requiredMatch
intencao✓ present✓ requiredMatch
agendamento_id✓ present✓ requiredMatch
escopo_autorizacao✕ not sent✓ requiredMissing
historico_conversa✓ present✓ optionalMatch
metadados_sessao✓ presentnot usedIgnored
escopo_autorizacao dropped out of the Front Desk → Billing contract. Billing rejected the request and fired two retries.
Context size12,4 KB · 3,2× the workflow average

Bloated context: the conversation history is duplicated across 4 agents in this workflow.

Recent runs on this handoff · last 24 h
tr_84921Rejected · scope missing12,4 KB420 ms09:42
tr_84918Rejected · scope missing12,1 KB404 ms09:38
tr_84911Completed8,8 KB362 ms09:31
Blast radius · 24 h
Runs affected128
Failure on this handoff17,8%
Wasted costR$ 184
Downstream agents2

Maestro analysis

Front Desk builds Billing's context from a template older than the authorization policy. It passes the patient state but not the scope. Billing validates every call and rejects 17,8% of the requests that come through this handoff.
Recommended fixAdd escopo_autorizacao to the Front Desk → Billing contract and validate it before every call. Projection: −17,8% failures, R$ 184/month saved.
Same pattern in other workflows
Triage → Paymentsescopo_autorizacao missing · 42 runs
Front Desk → Complaintshistorico_conversa missing · 18 runs

Incidents

Agent coordination failures that need attention

Open72 unassigned
Critical21.204 runs
Time to detect42 s−18 s this week
Time to resolve18 min−4 min this week
Wasted costR$ 612last 24 h
Incidents›INC-2481

Infinite delegation loop

Critical · ongoing

INC-2481 · Patient Care · Production · caught automatically by the loop detector

Started09:42:184 min ago
Runs affected842of 1.204 in the window
Wasted costR$ 1842,1M tokens
Agents involved2Supervisor · Research
Loop cycles17before shutdown
Detected in42 sfrom the 1st cycle

What happened

Supervisor delegated the same task to Research 17 times in a row. Research kept returning an answer with no completion state, so Supervisor read it as “still more research to do” and delegated again. The run burned 2,1M tokens before it hit the budget cap.

Failure path
SupervisorResearchSupervisorResearchSupervisorResearchLoop detected

Cycles 1 to 6 of 17 · each cycle costs R$ 10,80 and 124k tokens

Root cause analysis

Primary causeAmbiguous response contractResearch returns free text with no completion state.
Contributing factorIncomplete resultThe partial answer looked “in progress” to Supervisor.
TriggerRe-delegationSupervisor treats any answer without a status as “needs more research”.
OutcomeUnbounded delegationThere was no cycle cap on the Supervisor → Research handoff.

Maestro recommendation

High confidence · based on 4 similar incidents

Add an explicit completion state to Research's response contract and cap delegation depth in Supervisor.

Current contractresposta: texto: string
Proposed contractresposta: texto: string status: completo | incompleto | falhou max_delegacoes: 3
Expected impact: prevents about 842 wasted runs per week · saves R$ 184 per dayCreate fix
Incident timeline
Workflow run startedtr_84921
Research calleddelegation 1 of 17
Incomplete context returnedno completion state
Supervisor delegated againdelegation 2 of 17
50% of token budget1,05M tokens
Loop detector fired17 identical cycles
Workflow stopped842 runs affected
INC-2481 openedescalated as critical
On-call notifiedpatient care team

Fixes

Every contract or topology change Maestro proposed, tracked from Monitor to Block and verified on real runs

Tracked fixes

sorted by stage
How a fix moves forward
MonitorLogs every run that would break the contract. Changes nothing. At least 24 h before promoting.
WarnAttaches a warning to the trace and moves on. Promote once Monitor shows no false positives.
BlockStops the handoff before the agent is called. In production, it needs a second approval.

Roll back. Any fix goes back to the previous stage in one click.
Verification. A fix is verified when the metric it targets holds for 7 days straight on real runs.

Concept · AriadnePT