The Agent Control Plane for enterprise AI
See every agent. Control every cost. Govern every action.
The OpenTelemetry-native evidence layer above your existing observability stack — producing the audit artifacts your APM can’t, and triggering risk reassessment when agent behavior drifts.
The layer that starts where observability stops
Datadog, New Relic, Honeycomb, Grafana Tempo and Jaeger collect telemetry — that’s the data layer. A control plane interprets that telemetry against a governance policy and closes the compliance lifecycle: Article 12 record-keeping, Article 26 deployer accountability, risk reassessment on drift.
A company running 50 agents across Copilot Studio, CrewAI, LangGraph and custom pipelines needs one layer that knows what every agent did, what it cost, and whether it deviated from its registered behavior — before an audit or a runaway token bill makes the question urgent.
Sources: Nannini et al., arXiv:2604.04604 · Forrester ACP evaluation · OTel SemConv #159 #160 · By Henrique Veiga Curi, Founder, MeshAI Labs
Closes the feedback edge from Step 11 post-market drift detection back to Step 4risk reassessment — the loop most enterprises lack.
Four pillars of agent control
From zero-touch discovery through real-time policy enforcement — one layer, one inventory, one evidence trail.
Agent discovery & registry
Every agent that emits telemetry enters the registry automatically — registered or not. A living inventory of what is actually running.
- Zero-touch registration
- Unregistered agent flagging
- Health monitoring dashboard
- Framework-agnostic catalog
ML-powered anomaly detection
Four algorithms watch for cost spikes, reliability decay, behavioral drift and security threats on a live five-minute schedule.
- Z-score cost spike detection
- Error rate & latency decay alerts
- New model usage drift
- Dormant agent reactivation
Cost intelligence
Token-level spend attribution by team, project and agent — with guardrails that stop a runaway bill before finance finds it.
- Token-level spend attribution
- Budget guardrails & enforcement
- Model routing optimization
- ML-powered spend forecasting
Governance & compliance
Eight policy types enforced in real time at the proxy, immutable audit trails, and EU AI Act readiness scoring with HITL approval.
- Real-time policy enforcement (8 types)
- EU AI Act readiness score
- Human-in-the-loop approval queue
- Kill switch, ABAC, dependency graph
- Prompt injection & PII detection at proxy
Three forces are converging on the same quarter
Regulatory exposure
The EU AI Act ceiling for high-risk systems — up to €35M or 7% of global revenue. Deployers must show governance, audit trails and risk classification.
Agent sprawl
CyberArk counts 45+ non-human identities per employee. Most organizations cannot say which agents are running, what they touch, or what they cost.
Cost opacity
Token spend compounds while staying unattributed. Teams cannot see it, finance cannot forecast it, and nobody owns the budget line.
Purpose-built, not adapted
APM tools collect telemetry. Spreadsheets collect regret. Neither produces the artifact an auditor asks for.
“A control plane can’t govern what it can’t observe.”
Leslie Joseph · Forrester, 2026
Runs beside the stack you already bought
exporters:
otlp/meshai:
endpoint: ingest.meshai.dev:4317Observability platforms
In parallelAgent frameworks
11 nativeAI providers monitored
ProxyDeveloper tools
MCP serverLangSmith and LangFuse use proprietary callback APIs and don’t emit OpenTelemetry. MeshAI runs in parallel to them by design — we contribute to the OTel SemConv standard rather than chase per-vendor shims.
Yes. Like GDPR, the EU AI Act has extraterritorial reach. If your non-EU company’s AI system reaches the EU market or its outputs are used by EU residents, the same fines apply. Fines are based on global turnover: up to 7% worldwide.
Become a pilot partner
Free pilot for three months, white-glove onboarding, and a direct line to the founder. We’re shaping MeshAI with enterprise teams before GA.