The Agent Management platform is the heart of PHTN.ai — the operational framework that governs an agent's entire existence, from design and deployment to monitoring, optimization, and retirement.
One agent is easy. A fleet of agents touching production systems is an operations problem. Agent Management treats it that way.
Release agents into production with provisioned resources and token budgets. New agents are wired into existing multi-agent systems and integrated with your business systems through governed APIs.
Centralized tracking of accuracy, latency, and resource consumption — with real-time logging, error tracking, and performance analytics across every agent you run.
Version control for agents. A/B test agent versions in production and revert to a stable version instantly if an update misbehaves.
Policies for data privacy, security, and behavior are enforced at the platform level — access controls, audit logging, and guardrails that stop agents from violating business logic.
Performance data and user feedback drive continuous improvement — fine-tune agent behavior, retrain with new data, or retire underperformers.
Allocate, monitor, and enforce limits on compute, memory, and LLM spend across all managed agents — at every level of your organization.
Five versioned registries keep agents, models, prompts, tools, and knowledge governed and discoverable — the shared source of truth for the whole platform.
Registry
What it holds
Key Capabilities
Agents
Agent metadata, capabilities, agent-to-agent cards, skills, and versions
Lifecycle states from draft to archived; search by domain, capability, or ontology
Models
Model cards for 148 models across 14+ LLM providers
Health monitoring, circuit breakers, cost tracking, per-model permissions
Prompts
Prompt templates, variables, model bindings, and test cases
Version control, A/B testing, model-specific variants, evaluation scores
Tools (MCP)
Tool definitions, connectors, and vault-backed credentials
Governed onboarding, sandboxed execution, encrypted credentials
Knowledge Base
KB metadata, RAG pipelines, and document inventories
Six-level scoping, per-tenant vector isolation, live document sync
Supported by three further registries: ontologies (156+ concepts, 48 object types), pre-built RAG pipeline templates, and 53+ alert rules across the observability pillars.
Eight levels of multi-tenancy — from group and tenant down to domain, project, team, and individual agent — so business units, departments, and subsidiaries share a platform without sharing data.
Role-based access control ships with 13+ enterprise roles — including compliance officer and read-only auditor — and 106 fine-grained permissions validated within your organizational hierarchy. Segregation of duties is enforced by design: the people who build agents are not the people who approve production deployments.
Bring your identity provider — Microsoft Entra ID, Okta, or any OIDC, SAML, or OAuth 2.0 platform. IdP groups map automatically to platform roles, and joiner-mover-leaver changes sync via SCIM.
Approval gates are a first-class workflow node. Execution suspends at the gate; a reviewer approves, rejects, or modifies; execution resumes exactly where it left off.
Every invocation is logged — which agent, which tool, which parameters, what result, what cost — supporting access reviews and your audit calendar.
OpenTelemetry-native, with 61 fields captured per call and an 18-part correlation header tying every request together across services.
Layer
What is instrumented
Agent execution
Every invocation — inputs, outputs, model, tools, and evaluation scores
LLM calls
Provider, model, tokens, latency, cost, and quality on every call
Tool calls
Tool name, inputs and outputs, latency, and success or failure — with full audit context
Workflows
End-to-end timing, node order, and human-review wait times, streamed in real time
RAG queries
Query, retrieved chunks, relevance, faithfulness, and sources, evaluated per query
Telemetry exports to the SIEM you already run — Splunk, Sentinel, Elastic, or Chronicle.
We'll walk your tenancy model, roles, and observability in a live environment
Book a Demo
Or drop us a line at [email protected]
