Aether

Agency OS · Spec v1

Agents you can own.
Retainers you can bill.

Aether is the operating system for designing, grounding, evaluating, and installing AI agent systems inside the businesses you already build for — then keeping the monthly desk that watches them.

Open the studio Read the spec

Reverse-engineered from Voiceflow, Botpress, Sierra, Decagon, Fin, Copilot Studio, Dify, CrewAI, and the white-label reseller stack — then built for the gap they all share: agencies that ship the system, not a seat.

Elise · Meridian Dental

Warming the desk…

This is a real concierge on a seeded practice playbook. Try hours, a price, or an emergency.

Field notes, 2026

What the market already does — and where it leaves you hanging.

Enterprise CX is consolidating (Salesforce is buying Fin). Outcome-priced giants will not white-label for a twelve-client studio. Visual builders stop at conversation. Frameworks never become a demo. Aether is the missing middle.

Voiceflow

Conversation design

The most polished dialogue canvas in 2026. CX teams map turns, slots, and voice without code. Model-agnostic per step. Workflows plus Playbooks is the right split of deterministic vs generative.

Misses: Scoped to conversations, not crews or back-office work. White-label and custom domains sit behind enterprise. Export is a prototype, not a runtime you own. Per-editor pricing punishes agencies running 20 clients.

Botpress

Developer studio

State-machine thinking, multi-model routing, self-host option, no markup on LLM spend. Closest OSS cousin to a real runtime.

Misses: Steep learning curve. No agency partner billing. White-label is a developer project. Voice is missing. Multi-client operations are bolted on.

Microsoft Copilot Studio

Enterprise low-code

Native to M365, Dynamics, Teams. Generative orchestration plus classic topics. Channel coverage inside Microsoft.

Misses: Azure gravity. Transcripts have stored chain-of-thought in plaintext. Weak outside the Microsoft estate. You cannot drop the agent into a client WordPress site as a Lego block.

Dify

LLMOps / RAG

First-class knowledge bases, citations, versioning, API-first, self-host community edition.

Misses: Conversation design is thinner than Voiceflow. Agency retainers, white-label widgets, and outcome reporting are not the product.

Flowise / Langflow

Visual LangChain

Fast prototypes, huge node catalogs, true OSS licenses, familiar to engineers.

Misses: Experimentation platforms. Production governance, evals, handoff, and client-ready deploys are DIY.

n8n / Make / Zapier

Automation

The integration graph of the internet. Excellent when the agent must act in a thousand SaaS tools.

Misses: Not a conversation product. No brand voice, no widget, no CX handoff, no knowledge citations. Agencies still need a studio around it.

CrewAI / LangGraph / AutoGen

Code frameworks

Real multi-agent patterns, full control, no per-seat tax.

Misses: Months of engineering. 2026 research still shows almost no authorization confinement — a jailbroken specialist inherits the crew's credentials. Nothing an agency can demo to a dentist on Tuesday.

Sierra

Done-for-you CX

Agent-first, action-taking (refunds, flight changes), brand-faithful voice. Serves a large share of the Fortune 50. Outcome-priced. Guardrails are mature.

Misses: Custom contracts from ~$150K/year. You do not own the agent. No self-serve. Useless as a white-label retainer an agency resells. Switching cost is the business model.

Ten gaps

The work Aether is actually for.

01

Nobody built this for agencies

Voiceflow is for a CX team inside one company. Copilot Studio is for a Microsoft shop. Sierra is a managed service. An agency running 20 to 80 client retainers needs sub-accounts, white-label, cost-to-serve, changelogs, and a monthly report the client actually reads.

02

Lock-in is the product

OSSA's 2026 survey is blunt: MCP, A2A, ACP, AG-UI, and ANP move messages, but none define a portable agent. Switching overhead sits at 19–34%. Aether Spec is that missing contract: identity, graph, knowledge, tools, policy, evals, channels — in one JSON document you can leave with.

03

The stack is split across four tools

Conversation design lives in Voiceflow. Crews live in CrewAI. Knowledge lives in Dify. Channels live in n8n. Production systems need all four on one canvas, with a single export.

04

Handoff is a webhook in disguise

Most platforms escalate by opening a ticket. Real handoff is a context packet: summary, attempted actions, citations, urgency, SLA clock, and a warm transfer into the desk the business already runs.

05

Shipped without tests

Golden conversations, citation checks, forbidden-claim tests, and regression on prompt edits are still rare. Agencies get sued for a hallucinated price. Evals have to be a first-class object, not a lab notebook.

06

Vanity metrics

Message counts do not pay retainers. Appointments booked, quotes issued, tickets deflected, and after-hours coverage do. Observatory tracks outcomes, token cost, and SLA — the numbers you put on an invoice.

07

Agents hold the keys

2026 work on LangGraph, CrewAI, AutoGen, and MCP shows almost no confinement: a jailbroken specialist inherits the crew's credentials. Each Aether tool grant is scoped to an agent, not a bearer token for the whole workspace.

08

RAG is not a business

VentureBeat's 2026 pulse: 57% of enterprises have watched an agent be confidently wrong because the context was missing, not because the model failed. RAG answers what a document says. It does not remember the visitor, detect policy drift mid-session, or know what 'active customer' means in this company. Aether treats knowledge, policy, and memory as separate planes.

09

Agent washing

Gartner counted thousands of vendors marketing agentic capability and about 130 that actually have it. A chatbot with a retrieval box is not an agent. Aether makes the graph, tools, guards, and handoff visible so a client can see the difference.

10

The no-code ceiling

Visual builders handle conversation logic. Production then demands scenario tests, persistent memory, credential management, and an analytics pipeline they were never built to provide. Aether treats those as objects in the same spec.

Seven-layer lattice

One system, not four tools taped together.

  1. 01Channel fabricWeb widget, REST, Slack, email, MCP server, A2A Agent Card. One agent, every door.
  2. 02Session & memoryTurn state, conversation history, and facts the crew is allowed to remember.
  3. 03OrchestratorDeterministic graph walker plus a generative planner when the path is not scripted.
  4. 04Agent latticeConcierge, specialists, supervisor. Sequential, hierarchical, or swarm crews.
  5. 05Knowledge planeChunked sources, retrieval, citations. The model is not allowed to invent policy.
  6. 06Tools & policyLeast-privilege grants, PII masks, topic blocks, webhook and booking tools.
  7. 07ObservatoryTraces, evals, cost, outcomes, version changelog, retainer report.

Industry playbooks

Stand up a desk in an afternoon.

All playbooks

Lego block

Design here. Ship into the sites and systems you already run for them.

Aether Spec

One JSON document: agents, graph, knowledge, tools, policy, evals, channels. Move it between servers. Diff it in git.

White-label widget

A single script tag on the client site. Their brand, your runtime. No Aether chrome in the visitor’s face.

MCP · A2A · REST

Tools as an MCP server. An Agent Card for other agents. A plain POST for anything you already host.