Intelligent Orchestration NetworkPlatform

ION

The control plane for every AI agent you run: one inventory, one rulebook, one record and one bill, whoever built the agent.

About ION

The idea

Agents multiply. Control doesn't. ION sits above every agent you run, on every cloud, whoever built it, and gives you one inventory, one rulebook, one record and one bill.

Product tour

How ION works

  1. See the whole fleet

    The Overview shows four figures for the workspace: total agents, active agents, monthly spend and the latest security grade. Below them sit the recent agents with their platform and status, and the latest deployments. The workspace you are working in is always named at the top left.

    See the whole fleet
  2. Find agents already running

    ION syncs each platform connection and lists the agents it found there that are not in the inventory yet. Add to Agents brings one under ION's review and tracking, and Sync Now looks again.

    Find agents already running
  3. One inventory across clouds

    Every connected agent has a card with its platform, whether ION deployed it or it came from outside, and whether it is running or idle. Filter by platform or origin, search by name, or import an agent built elsewhere.

    One inventory across clouds
  4. Build a new agent

    AgentForge is the build studio. Name a product and follow it from research to a written brief, the agent workflow, a prototype page, build, test, deploy and live. Unfinished wizard runs keep their place and can be resumed.

    Build a new agent
  5. Run agents as one product

    A workflow joins several agents into one product with its own page. Each card shows the agents it runs, its status and when it was published; Version, Graph and Open let you revise it, see how it is wired, or use it. On the canvas, each run shows every step's input and output, and approval steps wait for a person.

    Run agents as one product
  6. Watch the cost

    Cost is broken down by platform and by category, such as inference, embeddings, tools and storage. Routing rules send requests to a chosen agent when a condition is met, each with a priority and an on or off switch.

    Watch the cost
The film · 0:41

See ION in motion

Sample data · synthetic demo
The problem

What ION catches

Case 01

Agents on every cloud, no single list

Teams have built agents on Azure, Google Cloud and AWS, and a vendor has added a few more. Each platform has its own console.

Red flag: Nobody can say how many agents the company runs, or where.

Case 02

A workflow that needs a tool you don't have

Someone describes a workflow and an AI drafts it. It looks complete, but one step relies on a tool the workspace has never connected.

Red flag: The step can't be bound to a real tool, so it shows as a gap and the workflow can't be switched on.

Case 03

A tool call that should be refused

An agent in a running workflow tries to use a tool. Its connection allows only a short list of tools.

Red flag: The tool isn't on the connection's allowed list, so the call is refused and the decision recorded.

Case 04

A bill with no owner

The monthly AI bill keeps rising. Spend is split across platforms and nobody owns the total.

Red flag: Cost by platform and category shows where the money goes, and a budget cap rule flags the overspend.

Stage by stage

How work moves through ION

  1. Idea

    Describe the workflow in plain words, or start from a research session in AgentForge.

  2. Plan

    An AI planner writes the steps, the trigger, and where decisions and human approvals sit. Gate 1: a person approves the plan, edits it or asks for a new one.

  3. Build

    Each step is bound only to agents and tools that really exist in the workspace, so nothing is invented. The draft is checked and repaired up to three times.

  4. Gaps

    Gate 2: a checklist of what can't run yet, such as an agent still being built, a missing tool, a setting to fill in or a webhook secret. The workflow can't be switched on until it is clear.

  5. Test

    Gate 3: each step is probed, then one supervised test run goes through the real engine.

  6. Live

    Once the test passes, the workflow can be switched on. The live canvas shows each step's input and output, and approval steps wait for a person.

Ground rules

Promises it always keeps

  • No approval, no deploy

    The deploy step refuses to start until a reviewer has approved the agent. That gate can't be skipped.

  • A no always wins

    When policy rules disagree, the refusal stands. An allow written for one agent never overrides a deny.

  • When in doubt, refuse

    If a connection's allowed list can't be read, or the policy check is unavailable, the tool call is refused, not let through.

  • Nothing goes live untested

    A generated workflow stays switched off until every gap is cleared and a supervised test run passes.

  • Every change is on record

    Each lifecycle step writes an audit entry, and policy decisions go into a log where each entry is linked to the one before.

  • Tenants never share data

    Every query is filtered by tenant, and PostgreSQL row-level security enforces the same boundary in the database itself.

Under the hood

What's inside ION

Registry

The inventory: every agent ION manages, with each version of its definition, and the agents found on connected platforms.

AgentForge

The build studio. A guided AI wizard takes an agent from an idea to a platform-specific package, researching the tools it needs on the way.

Workflows

Joins agents, tools, decisions and approvals on a canvas, runs them as one product with its own page, and shows each run step by step.

Governance

Runs the lifecycle and the approval gate, and checks each agent against 10 rules across security, compliance, cost, access and resilience to give a grade from A+ to F.

Connectors

Connects to each cloud, finds the agents already running there, and deploys new ones.

Observability

Collects agent telemetry, works out cost by platform and category, and scores each agent on reliability, latency, cost and throughput (AgentRank).

By design

What it deliberately won't do

  • It won't: Deploy an agent nobody approved

    The deploy step checks the agent's status first and stops unless a reviewer has said yes.

  • It won't: Invent tools for a workflow

    A generated workflow can only use agents and tools that already exist in the workspace. Anything missing shows up as a gap to fill.

  • It won't: Switch on an untested workflow

    Activation stays blocked until the gap checklist is clear and a test run passes.

  • It won't: Let a tool call through when unsure

    A call is refused when its connection's allowed list can't be read or the policy check is down.

  • It won't: Mix one customer's data with another's

    Each tenant's data is fenced off down to the database rows.

  • It won't: Tie an agent to one platform

    The same agent definition can go to Azure, Google Cloud or AWS, or run in ION's own sandbox, without a rebuild.

Technical details

ION is a set of .NET 10 services behind one gateway: Registry (the agent inventory and definition versions), Forge (the build wizard), Governance (lifecycle, rules and the approval gate), Connectors (platform connections, discovery and deploys), Observability (telemetry, cost and AgentRank), an Orchestrator (task queue and workers) and AgentMesh (sessions between agents). Traffic enters through a YARP gateway that checks each sign-in token, and the services talk through MassTransit over Redpanda, a Kafka-compatible message broker. Data sits in PostgreSQL 17, one schema per service. Tenant isolation has three layers: request middleware, a tenant filter on every query, and PostgreSQL row-level security. Sign-in uses Zitadel (OpenID Connect) and secrets live in Infisical. Policy checks run on Microsoft's Agent Governance Toolkit for .NET, set so that a deny always overrides an allow, and governance decisions are written to a hash-chained audit log. Agent telemetry arrives over OpenTelemetry (OTLP), is buffered in Redpanda and stored by Observability; it can also be forwarded to the customer's own Grafana Cloud. The dashboard is React 19 with Vite, TypeScript, Tailwind CSS 4 and React Flow, with live updates over SignalR. It runs in containers on K3s with Helm and ArgoCD. An MCP server lets developers build and wire agents from Claude Code with the ION plugin. Some labels inside the product still use its working name, Nexus.

Target audience

CIOs, CTOs and heads of AI governance who run AI agents across several clouds and vendors. Day to day it is used by platform engineers, DevOps and AI team admins who check builds, approve deploys, fix connections and manage who can reach each workspace, and by members who work with their own workspace's agents. Reviewers who approve agents and finance owners who watch AI spend use it too.

Questions

Which clouds can it deploy to?

Azure, Google Cloud or AWS AgentCore, or ION's own sandbox. The same agent definition can go to more than one without a rebuild. ION also finds agents already running on connected platforms and adds them to one inventory.

Can an agent go live without approval?

No. Agents move through Draft, Review, Approved, Deployed and Retired in that order. The deploy step refuses to start until a reviewer has approved the agent, and every change is logged.

What are the three gates?

When ION builds a workflow from your description, a person first approves the plan. Then a checklist of gaps, such as a missing tool, a setting to fill in or an agent still being built, must be cleared. Last, every step is probed and one supervised test run must pass. Only then can the workflow be switched on.

How is an agent graded?

ION runs 10 rules on each agent: least privilege, OAuth scopes, secret rotation, data residency, audit log retention, budget caps, idle agents, rate limits, role setup and fallback chains. Results are weighted by severity and turned into a grade from A+ to F, overall and for each area.

What does "a no always wins" mean?

When two policy rules disagree about a tool call, the refusal stands. If a connection's allowed list of tools can't be read, the call is refused rather than let through. Each decision is recorded.

Do we have to write code to build an agent?

No. The guided AI wizard in the browser takes you from an idea to a deployable agent, and workflows can be described in plain words. Developers can instead push a ready-made agent definition from a laptop or build pipeline, or ask for an agent from Claude Code with the ION plugin. Every route goes through the same review.

Can we bring in agents built elsewhere?

Yes. Agents found on a connected platform can be added to the inventory, and a package built elsewhere can be imported in five steps: upload, checks, platform, build and deploy. A draft is saved at each step, so you can stop and come back.

Is our data kept apart from other customers?

Yes. Each tenant's data is kept apart at three levels, down to row-level security in the database. Inside a tenant, workspaces group agents, connections and people, with Member and Admin access. Connector credentials are encrypted.

Has it been security tested?

A VAPT (vulnerability assessment and penetration testing) certificate is available from BISTEC on request.

What does it cost?

Pricing, service levels and terms are provided by BISTEC on request.

Book a demo

See any of the nine running on your own use case.

A 30-minute walkthrough with the team that built it.

Email us
hello@bistecglobal.com

Offices in Sydney and Colombo.

bistecglobal.com
© 2026 BISTEC Global · Intelligent design. Engineered for adaptability.Demos on this page use sample data.