--- title: Agent Platform — Enterprise AI Orchestration | Zentavor description: The orchestration layer Zentavor uses to build, train and run enterprise AI agents: multi-agent by default, auditable by construction, on-prem ready. canonical: https://zentavor.com/agent-platform.html --- # Agent Platform — Enterprise AI Orchestration | Zentavor > The orchestration layer Zentavor uses to build, train and run enterprise AI agents: multi-agent by default, auditable by construction, on-prem ready. Products / Agent Platform # Agents that ship to production Agent Platform is the orchestration layer Zentavor uses to build, train and run AI agents inside enterprises. Multi-agent by default, auditable by construction, on-premise when the regulator says so Request a demo How it works → Runs where your data lives: Azure, AWS, GCP or bare-metal. 45% shorter time to production −30% lower operational workload 37% more predictable delivery How Agent Platform works ## One request. Six stages. One platform Follow a single request as it travels through the stack: from the channel it arrives in, to the systems it changes, under governance the whole way. Scroll to watch each stage light up Surfaces ### Work begins where people already are A request arrives in the tools your teams already live in: Teams, Slack, the web, an IDE, voice or email. Agent Platform meets it there, with no new app to adopt. MS Teams Slack Web IDE Voice Email Orchestration ### The core plans the work Agent Platform core decomposes a business goal into reproducible steps: tools, retries and branching, all observable. A LangGraph runtime carries approval and escalation flows from the first step. LangGraph runtime Approval & escalation A2A protocol Reasoning ### Agents do the thinking Boring tasks run autonomously; hard tasks run with a co-pilot. Each step picks the right model for the job: Claude 4, a GPT-class model, a self-hosted model or a specialist one. Claude 4 GPT-class Self-hosted Specialist models Retrieval ### Grounded in your own data Agents analyze documents and data (OCR, RAG over corporate stores, SQL over the warehouse) inside a single context window. Vector and hybrid search, schema linking, MCP tools. Vector DB Hybrid SQL+semantic Schema linking MCP tools Action ### It acts in your systems Agent Platform writes back to Salesforce, SAP and ServiceNow, triggers ETL and refreshes BI, with scopes enforced on every call. CRM/ERP write-backs ETL triggers BI Custom tools Trust ### Auditable by construction Every step carries approvals, scopes and an audit trail. PII detection, a reviewer UI and cost & SLA telemetry: governance your security team will recognize. Audit log PII detection Reviewer UI Cost & SLA telemetry Agent Platform core Orchestrator Policy guardrails Governance & approvals Support agent Customer support Sales agent Lead & deal management Ops agent Operations & monitoring Tools & APIs Integrations Demo · pick a scenario ## Watch Agent Platform decompose a real task Click a scenario and see the agent break the job into tools, retrievals and write-backs. This walks through the recipe the agent follows Banking Triage & resolve a support ticket AiDoc Parse a supplier invoice end-to-end TextToSQL Answer a business question with SQL Data Ops Diagnose an ETL incident Trading Trader copilot · summarize incidents agent.run #a3f91c Banking · support ticket resolved · 4.7s Customer · #4471 Classify classifier · 6 classes 0.4s account card billing fraud loan general Retrieve rag.kb · scope=billing 1.1s KB-118 · Card replacement policy 0.41 KB-204 · Double-charge refund procedure 0.91 KB-077 · Statement dispute handling 0.58 Draft llm.claude-4 · grounded 2.3s Agent Platform → customer Customer 4.7s end-to-end · servicenow.update · audit-log + cost ✓ agent.run #b7e2 AiDoc · supplier invoice booked · 4.2s Invoice · OCR invoice_4471.pdf Extracted vendor Globex Ltd PO PO-4471 net €8,400.00 VAT 20% €1,680.00 due Net 30 OCR aidoc.ocr · layout-aware 1.8s Match erp.search · open POs 0.9s PO-4310 · Globex Ltd · €6,200 0.39 PO-4471 · Globex Ltd · €10,080 0.97 PO-4502 · Initech · €3,540 0.12 Verify rules.tax 0.5s VAT 20% valid within contract amount ≤ PO Approve workflow.route 0.3s Output · booked to SAP journal#JV-88123 payable €10,080.00 due 09 Jul · Net 30 GL 5100 4.2s end-to-end · accounting notified · audit-log + cost ✓ agent.run #c4a1 TextToSQL · business question answered · 3.4s Question · analyst Understand llm · intent 0.6s metric net_revenue by region period 2026-Q1 Link schema schema.link · 6 tables 0.4s customers orders order_items regions products invoices Generate sql.gen · grounded 1.3s Validate sql.dry-run 0.8s syntax ok 1,240 rows scanned types ok Output · result | Region | Net revenue | QoQ | EMEA | €4.2M | +18% | Americas | €3.1M | +6% | APAC | €1.8M | +11% 3.4s end-to-end · query validated · audit-log + cost ✓ agent.run #d9f3 Data Ops · ETL incident recovered · 4.8s ⚠ Pipeline failed: orders_etl · stage load 02:14 UTC · severity high · batch aborted Detect monitor · orchestrator 0.2s Diagnose logs.search · orders_etl 2.1s 02:14:02 INFO extract · pulled 12,402 rows from source 02:14:07 INFO transform · 12,402 → 12,402 (0 dropped) 02:14:11 INFO load · connection to warehouse ok 02:14:12 ERROR load · IntegrityError: null value in column "customer_id" violates not-null constraint 02:14:12 DEBUG failing row → id=8231, order_no=SO-77419, customer_id=NULL 02:14:12 ERROR load · batch aborted (rollback complete) Runbook kb.runbook 0.5s RB-04 · Source connection reset 0.22 RB-12 · Null-key backfill & safe retry 0.95 RB-19 · Schema drift reconciliation 0.31 Act pipeline.retry · safe 1.4s Extract→Transform→Load Output · recovered pipeline orders_etl status green rows 12,403 ticket#INC-2291 4.8s end-to-end · on-call notified · audit-log + cost ✓ agent.run #e1c7 Trading · incident brief delivered · 6.1s Request · trader Gather rag.incidents · this week 1.2s Settlement delay · EUR book Margin-call latency Limit breach · FX desk Data-feed gap · Rates Settlement delay · UST Limit breach · Equity Stale quote · Credit Settlement delay · JGB Cluster embed · by theme 0.9s Settlement ×3 Limit breaches ×2 Data / feed ×2 Summarize llm · grounded 2.6s ↳ 23 incidents · 4 desks Compare sql · vs 4-wk avg 1.0s 23 incidents · −22% vs trailing 4-week average Output · brief delivered incidents 23 vs avg−22% top theme Settlement delivered PDF · trader 6.1s end-to-end · audit-log + cost ✓ Built for delivery ## Why teams pick Agent Platform instead of assembling LangGraph from scratch ### Assemble it yourself Glue LangGraph, queues and tools together by hand. Build approval, escalation and human-in-the-loop from scratch. Add audit logging and cost tracking after the fact. Re-solve deployment and security for every new use case. Months of platform work before the first agent ships. ### Agent Platform Multi-agent orchestration with shared context out of the box. Approval, escalation and human-in-the-loop as primitives. Audit log and cost telemetry on every step, by construction. On-premise ready: Azure, AWS, GCP or bare-metal, one security model. Playbooks for six domains. From pilot to production with one partner. Governance ## Auditable by construction Every step an agent takes is scoped, logged and reversible. It is the same trail your security team expects from any production system, not a black box Audit trail run #4815 · support.triage - retrieve kb:read system ok - draft llm:claude-4 PII masked redacted - decide policy:resolve a.petrov approved - write-back servicenow:update scoped token ok latency 4.7s cost $0.012 SLA within ### Scoped on every call Agents act through least-privilege tokens, never your standing credentials. ### PII detected & masked Sensitive fields are caught and redacted before they ever reach a model. ### Human in the loop High-stakes steps wait for a named approver, with the reason logged. Runs where your data lives Azure AWS GCP Bare-metal On-prem / VPC See it live ## See Agent Platform in your environment A 45-minute technical session with a delivery lead, using your IT landscape and governance model as input Request a demo Explore AiDoc →