Platform
Claudia's fleet
The public operating model defines 20 agents in 9 classes, with responsibilities and governance levels for each.
20 agents · 9 classes
- AdvisoryConciergeStudio
- Service DeskTriageDispatchResolutionSelf-ServiceChange
- Endpoint / Device OpsNOCEndpoint OpsBDR
- Security (SOC)SOC
- Client Advisory (vCIO)vCIOAccountBilling OpsVendor Mgmt
- GRC / ComplianceGRC
- Analytics / InsightsWatchdogInsights
- OrchestrationOrchestrator
- Knowledge / DocumentationDocumentation
Claudia is invite-only for a limited supervised launch. Starting with Autotask + Datto RMM and ConnectWise PSA + ConnectWise Automate.
Claudia’s AI agents resolve tickets end to end: they ask the right questions, get your technician’s approval, fix the problem, document it and close the ticket. The catalogue below describes the broader model; each pilot's capabilities and endpoint scope are agreed with your team.
20
AI agent roles
77
Agent skills
157
Workflows
15,000+
Tools across 70 connectors
The model's five service-desk roles
The five service-desk agents of Claudia's fleet
Each represents a stage of the model's ticket lifecycle. The screens illustrate those responsibilities; launch capabilities are scoped separately.
| Stage | Agent | What it hands on | From |
|---|---|---|---|
| 01 | AI Client ConciergeDeflect · Collect | Deflected, or the missing detail collected | Pro |
| 02 | AI Intake & Triage AgentVerify · Classify · Prioritise | Matched to its record, prioritised on the SLA | Starter |
| 03 | AI Dispatch & Scheduling AgentRoute · Pre-flight | Routed on skill, load and availability | Pro |
| 04 | AI Autonomous Remediation AgentResolve · Verify · Close | Runbook run, closure notes written | Starter |
| 05 | AI Service Desk WatchdogWatch · Alert | SLA clock watched, a human alerted in time | Elite |
The illustrations cover 4 of the 5 service-desk stages. Launch work stays under human supervision within the agreed pilot scope. What setting this up involves.
The whole fleet
All 20, by class
Every agent in the model's words, with its class's governance ceiling. These model definitions do not establish runtime approval or execution evidence. The current launch is supervised.
Knowledge / Documentation
1 agent · model ceiling: L2 Supervised
“Every write is a versioned proposal the store can revert, but a wrong runbook is acted upon by a technician — so unattended publication waits for the store's review telemetry. Reads everything a client's documentation set holds; never a secret value.”
- DocumentationL1 CopilotCreates and maintains the MSP's documentation layer — KB articles curated from resolved tickets, runbooks, configuration and flexible-asset facts captured from discovery, policies routed for acknowledgment — and keeps it honest: coverage against the reference's documentation types, freshness against each type's lifecycle, drift against the configurations it describes, duplicates retired by archiving. Every write is a versioned proposal; secret values never.
Service Desk
5 agents · model ceiling: L2 Supervised
“Per-ticket scope on high-volume, mostly reversible work. L1 required a human to approve every single ticket classification, which is not defensible for triage. L3 only for read-only deflection.”
- TriageL1 CopilotIngests raw ticket descriptions, emails, and calls: executes caller identity verification (Gap 02), categorizes issue to canonical records, sets priority, detects VIP sentiment (Gap 09), and clusters duplicate incidents (Gap 03).
- DispatchL1 CopilotEvaluates technician skill rubrics, tier fit, pod load, and calendar availability to recommend assignments, pre-flight tool access, and orchestrate swarming handoffs (Gap 05) for stalled tickets.
- ResolutionL1 CopilotGuides technicians through verified step-by-step resolution checklists, executes automated runbooks for tier-1 incidents, validates First-Contact Resolution (Gap 06), and generates comprehensive closure documentation.
- Self-ServiceL1 CopilotAnswers end-user questions in plain language, flags self-resolve eligibility, and lists prerequisites — never internal jargon.
- ChangeL1 CopilotAssesses Change Requests for collisions, evaluates CMDB blast radius, pre-approves low-risk standard changes, validates rollback and verification test plans for the Change Advisory Board (CAB), and oversees 30-day onboarding delivery checklists and project-phase gates.
Orchestration
1 agent · model ceiling: L3 Guardrailed
“Routes work and never executes a worker's action. Its failure mode is misrouting, and the receiving worker's own gate still fires.”
- OrchestratorL2 SupervisedThe supervisor over the specialist fleet: classifies inbound work, routes each item to the right specialist squad, monitors fleet health, and escalates ambiguous or cross-domain work without executing worker actions directly.
Endpoint / Device Ops
3 agents · model ceiling: L3 Guardrailed
“Gated at the ACTION grain: the RMM script autonomy triad keeps approval-required and human-only scripts human-gated whatever the agent's level. Highest operational blast radius.”
- NOCL2 SupervisedOwns network infrastructure and telemetry for the specialist fleet: triages monitoring telemetry (SNMP, syslog, cloud metrics), suppresses alert noise by ≥70%, correlates signals against the CMDB topology (Gap 07), and determines blast radius before spawning incidents.
- Endpoint OpsL1 CopilotOwns endpoints, scripts and patching for the specialist fleet: governs fleet-wide endpoint script execution (the governed RMM script catalog), patch compliance, and scheduled maintenance windows, enforcing strict technician consent gates (Gap 10) on disruptive or reboot actions.
- BDRL1 CopilotProtects client data integrity: audits nightly backup jobs across Datto/Veeam, evaluates automated boot screenshots, checks cloud air-gap immutability, and coordinates quarterly test restores.
Security (SOC)
1 agent · model ceiling: L2 Supervised
“HELD DELIBERATELY at L2. A wrong containment action isolates a healthy production host. L3 only ever for detection and enrichment, never for containment.”
- SOCL2 SupervisedHigh-speed security incident response across endpoints and cloud identities: isolates compromised hosts, revokes compromised M365 sessions, triages phishing reports, and maintains chain-of-custody audit logs.
Client Advisory (vCIO)
4 agents · model ceiling: L2 Supervised
“Per-client scope with client-visible output. Acts within scope, but anything that reaches the customer keeps approval. NEVER L3.”
- vCIOL1 CopilotDelivers strategic technology consulting: analyzes asset lifecycles, end-of-life operating systems, and tech debt to generate automated Quarterly Business Reviews (QBRs), multi-year roadmaps, and capital budgets.
- AccountL1 CopilotManages the commercial client lifecycle: monitors agreement contract expiration dates, tracks license counts and true-ups, and detects upsell opportunities.
- Billing OpsL1 CopilotOperates Gap 12: audits unbilled technician time, books AI labor against PSA contracts with proper credit models, calculates Effective Hourly Rate (EHR), and pre-flights invoice batches to eliminate margin leakage.
- Vendor MgmtL1 CopilotOperates Gap 11: tracks external support cases opened with Microsoft, Cisco, Dell, and ISPs, enforces vendor Operational Level Agreements (OLAs), and triggers automated escalation cadences on stalled vendor tickets.
GRC / Compliance
1 agent · model ceiling: L2 Supervised
“Evidence gathering is read-heavy and repetitive, which is what L2 is for. Attestation submission is rare and high-stakes and stays approval-gated.”
- GRCL1 CopilotMaintains continuous compliance readiness: audits tenant controls against CIS Controls, NIST CSF, HIPAA, and CMMC, tracks security baseline drift, updates risk registers, and compiles audit evidence binders.
Analytics / Insights
2 agents · model ceiling: L3 Guardrailed
“Tenant-scope reads on a schedule with no write surface. Produces recommendations and monitor-detect flags without mutating technical configurations.”
- WatchdogL1 CopilotMonitors active ticket queues 24/7: tracks real-time SLA pause/resume countdown clocks (Gap 04), enforces three-strike customer communication cadences (Gap 08), and alerts managers to breach risks and stalled tickets.
- InsightsL0 AdvisoryAnalyzes ticket patterns, costs, volumes and avoidability across the reference; recommends proactive levers; detects recurring incident clusters (Gap 03) and authors problem records with root-cause hypotheses — the fleet's ITIL problem-management owner.
Advisory
2 agents · model ceiling: L3 Guardrailed
“Drafts only — executes nothing outside the drafting loop, so unattended operation still produces a draft a human reads. NEVER L4: a draft nobody reviews is a publication.”
- ConciergeL1 CopilotGuides a visitor through the reference: pulls their PSA config, builds the 9-axis profile, runs the diagnostic, navigates, and assembles QBRs / client emails on demand.
- StudioL1 CopilotThe operating model's build studio: authors and validates n8n automation workflows, generates PowerShell/Bash RMM scripts with safety parameters, extends taxonomy categories, and evaluates agent test suites and OML maturity.
Agent catalogue and class ceilings from The Agentic MSP Operating Model, version 4.361.0, derived on 2026-09-22.
The e-book · 25 pages · PDF
The Agentic MSP Operating Model, explained
The public model the agents reason from, as a field guide you can map your own desk to.
Graduated autonomy
Every agent starts supervised. Your MSP decides how far it goes.
The current launch is supervised. The broader operating model describes graduated autonomy; any later expansion requires an agreed scope and evidence from the actual workflow.
- L0Claudia today
Advisory
Recommends only; a human does everything.
To reach L1Advisory recommendations are accepted (acted on) ≥80% of the time over a meaningful sample — the suggestions are trustworthy enough to draft from.
- L1Claudia today
Copilot
Drafts/acts on request; a human approves each step.
To reach L2~200+ copilot actions approved with ≤5% correction rate over ~4 weeks — the human is rubber-stamping, not rewriting.
- L2Claudia today
Supervised
Acts within scope; a human supervises + can intervene.
To reach L3≥6 weeks supervised with no intervention-worthy error and the quality metric at/above threshold — scope is well-bounded and the failure modes are understood.
- L3Permitted, on evidence
Guardrailed
Acts autonomously inside hard guardrails + budgets.
To reach L4Sustained in-guardrail operation under outcome (not step) review, with no guardrail breach and stable quality — humans review results, not actions.
- L4Not permitted
Autonomous
Runs end-to-end; humans review outcomes, not steps.
0 of 20
Agents cleared to act without a human
L2
Highest level any agent runs at today
The levels and the rungs are the operating model's, verbatim; the rung figures illustrate the shape of the bar rather than measured telemetry. Every agent starts supervised on every MSP, your team raises the level on evidence, and the highest level any class of agent is permitted today is L3. No agent is cleared to act past the checkpoint your MSP sets — the two figures above are derived from the operating model, not asserted here.
Want to take Claudia for a spin in your IT environment?
Claudia is invite-only for a limited supervised launch. Invited MSPs receive a 30-day free trial in an agreed test or pilot environment — no card, and no onboarding fee. Scope is confirmed before access.