Skip to main content
An AI Harness lets your Flowyte agent hand a live conversation to your own operator agent. When a chat or SMS conversation meets a rule you set, Flowyte opens a durable escalation and notifies your harness (Hermes, OpenClaw, or any runtime you build). Your harness claims the conversation and talks back and forth with the customer in the same thread, then resolves it, returns control to the Flowyte AI, or asks for a human. Flowyte owns the conversation record and keeps enforcing compliance; your harness is a participant. It reaches the customer in their original thread — a chat widget shows a “specialist has joined” divider; an SMS thread continues as a normal text.

​
Set it up

1

Connect your harness

In Integrations, connect your AI Harness provider (hermes or openclaw). The escalation surface stays hidden until at least one harness is connected.
2

Create a connector key

Mint a scoped API key for your harness in the dashboard’s Developer page, granting only the escalations:* scopes it needs (see the connector reference). A connector key can work conversations but can’t edit your destinations or agents.
3

Add a destination

Register where conversations go: POST /escalation-destinations with your harness’s endpoint_url and the channels it handles (chat, sms, or both). The response returns a signing secret (whsec_…) once — store it; your harness verifies webhook signatures with it. The destination starts unverified.
4

Run the handshake test

POST /escalation-destinations/{id}/test opens a sandbox escalation and calls your endpoint. Your harness must receive → claim → reply → resolve within the window. Only a green run flips the destination to active; the response reports a step-by-step ledger with a hint on failure.
5

Set the agent's escalation policy

PUT /agents/{id}/escalation-policy sets an ordered list of rules — when to escalate and where to route. Rules match on deterministic conditions (an explicit request for a person, a human request, business hours, channel, a failed tool or knowledge lookup) and route to an active destination, an email fallback, or a queue.
6

Publish the agent

The policy edits the agent draft. Publish so live phone and chat traffic use it.
The policy is per agent — two agents in one organization can route to entirely different harness destinations. Destinations are shared organization-level infrastructure; the policy decides which agent uses which.

​
What runs a conversation

Once a destination is active and a policy rule matches, a live conversation follows this loop:
  1. Flowyte opens an escalation and sends your harness a signed webhook naming the channel and a short reason.
  2. Your harness claims the escalation, reads the handoff context (a transcript window plus the escalation reason and any verified identifiers), and starts replying.
  3. Each reply reaches the customer in their original thread; each customer message is forwarded to your harness. While your harness owns the thread, the Flowyte AI stays silent.
  4. Your harness ends the conversation — resolve (done), return (hand back to the AI, which resumes on the next customer message), or request a human (routes to your human fallback).
Building the harness side? See the connector reference for the signed-webhook, claim, reply, and resolve contract.

​
Rules to know

SMS is customer-initiated only. Your harness can reply inside an SMS conversation the customer started, but it cannot open one. Outbound initiation is coming soon, tied to Flowyte’s outbound campaign rules.
  • SMS requires registration. Enabling the sms channel on a destination requires your organization’s completed A2P 10DLC (TCR) registration — the same registration your inbound SMS number already needs. See SMS.
  • Plan. Harness conversations are included on Starter and above. Setting up destinations and running the handshake test work on any plan, so you can wire everything up before you upgrade.
  • Compliance stays with Flowyte. Opt-out, quiet hours, and consent are enforced on your harness’s replies exactly as they are on the AI’s — a reply to an opted-out recipient is suppressed, and a quiet-hours send is held.
  • One owner per thread. A conversation has at most one non-AI owner at a time, so a harness and a human seat can never both hold the same thread.

​
In the API

The escalation endpoints are in the API Reference. Configuration endpoints (destinations, policy) work on any plan; the runtime endpoints are what your connector uses. Authenticate every call with Authorization: Bearer flowyte_sk_…. You can also drive the same surface over the MCP gateway — the escalation tools map one-to-one onto these endpoints. Next: the connector reference for building the harness side.