Measure
Agentic systems
Per-agent flows, provider calls, message rates and the agent-to-agent delegation graph across hosts. LangGraph, CrewAI, AutoGen or your own agents, in one stream.
Research & Education · Free educational license
Ground truth for the agentic era.
One consistent view of every agent in your multi-agent system, across frameworks and hosts, including the ones you didn't write. Payload-blind, with nothing instrumented.
What it is
A measurement instrument, not a platform.
Maya records what each agent reaches, who it delegates to, and where it strays from what you declared. What that means is yours to decide.
Study
Agent safety and control
Declare an agent's authorized envelope and see Maya flag when observed behavior contradicts it: misbehavior, containment from outside the agent, a valid credential acting out of role.
Teach
Courses and labs
A reproducible multi-agent environment where delegation graphs, deviations and containment become hands-on exercises, without every student instrumenting every agent.
What makes it different
The view app-layer tracing can't give you.
Payload-blind by construction.
Never terminates TLS, never decrypts. Maya reads envelope metadata (who, where, how much), not your content.
Per-agent, not per-host.
Identity is derived from the host, per agent, with no cooperation from the agent. It sees the black-box agents you didn't write.
Governs only what you declare.
Maya watches the agents you name and nothing else. No silent surveillance of unrelated workloads.
Emit-first and programmable.
Structured deviation telemetry, exportable as OTLP or JSON, routed to your own judge: a notebook, an LLM, Open Policy Agent or your SIEM.
How it works
Four steps, on your own infrastructure.
Two packages: one on each agent host, and the data plane appliance that governs their traffic. Your data never leaves your environment.
Discovery
A short call to map your host inventory and what you want to study or teach.
Provision
We hand you the two packages, recommended compute specs and a runbook.
Declare
One line in each agent's existing deployment config names it. No code changes.
Observe
Read per-agent telemetry, map delegation graphs, flag deviations, plug in your own judge.
Where Maya is today
Active beta. We tell you what's proven.
Today
What we lead with.
Per-agent observation, delegation-graph mapping, and declared-versus-observed deviation telemetry. Programmable enforcement is available and evolving.
We are direct about what is proven and what is in progress, and a design partnership is where that conversation belongs.
Your setup
If you run the hosts, Maya runs there.
- Runs on Linux hosts or Kubernetes nodes you control: bare VMs or your own cluster.
- Deploys in your own lab or cloud; your agent data never leaves your environment.
- Not a fit for fully managed serverless agent runtimes where you don't control the node, such as Fargate or Snowpark.
Design partners
Bring ground truth to your agent research.
Each term we partner with a limited number of labs, universities and educational programs: the full platform under a free educational license, a runbook and hands-on onboarding. In return, candid feedback, and case studies or citations only if and how you're comfortable.
Email research@mayagentic.com and tell us about your lab.
Maya Research Edition · Educational license, free of cost
Free · two weeks · your environment
Get an Agent Exposure Assessment.
See which agents are operating, where they run, and what they communicate with.