Agents

Register a webhook agent by pasting a curl command, or pick one from the Agent Registry. Registered agents become tools Flux can call, and you can mention them by name.

Agents extend what Flux can do with capabilities you own. Two kinds appear on this page:

Webhook agents. Anything callable over HTTP. You register it by pasting a curl command and it becomes a chat tool.
A2A-protocol agents. Discovered automatically from the configured A2A services, and reached agent-to-agent over the A2A client.

The Agents page showing registered agent cards with POST and A2A badges, host, secret and header counts, last-called times, body schema fields, and an Exposed to coordinator toggle

Registering an agent

Add agent opens a dialog with two tabs: pick from the Agent Registry, or paste a Curl webhook.

From the Agent Registry

Agents published to Agent Fabric appear here. Search by name, capability, or publisher, tick the ones you want, and press Done.

The Register an agent dialog on the Agent Registry tab, showing a searchable list of agents with version numbers, approval status badges, capability counts, hosts, publishers and capability chips

Each row tells you enough to choose without leaving the dialog:

On the row	Meaning
Name and version	Which agent, and which published version
Approved / Pending approval	Its governance state in the registry
N capabilities	How many distinct things it exposes
Host	Where calls go
Publisher	Who published it, for example your own organization or a public A2A registry
Capability chips	The named capabilities, for example `execute` or `film_video_at_location`

Read the publisher and the approval state. A registry can carry agents scanned from public A2A directories alongside your own. An agent marked pending approval has not been through your governance process, and adding it here gives the coordinator a route to a third party's endpoint.

Imported agents are labelled Imported from the Agent Registry on their card, and Refresh registry re-reads the list after someone publishes a new one. If the registry is empty, nothing has been registered in Agent Fabric yet, which is a platform-side task rather than a chat-side one.

From a curl command

The fast path for anything you host yourself. Paste a working call and the fields are parsed out of it.

The Register an agent dialog on the Curl webhook tab, with a pasted curl command showing the endpoint, content-type and API key headers, and a JSON body

curl -X POST "https://your-service.example.com/webhook/marketing-thesis/default" \
  -H "Content-Type: application/json" \
  -H "X-API-Key: <token>" \
  -d '{"focus_areas":["AI in drug discovery"],"partner_name":"Acme Labs"}'

From that it derives the method, the URL, the headers, which of them are secrets, and a body schema from the JSON keys. Press Register agent, then give it two things of your own:

Field	Why it matters
Name	How you mention it in chat, for example `@sop-agent`
Description	What this agent does, surfaced to the coordinator

The description is the routing signal. Flux decides whether to call an agent from its description alone, so write what it is for, not what it is built with. "Answers questions about standard operating procedures from the SOP corpus" gets called. "SOP microservice" does not.

What a card tells you

On the card	Meaning
POST and A2A badges	How the agent is called
Host	Where the call goes
Secrets and headers count	How many header values are stored, and how many are secret
Last called	The last time Flux used it, or Not yet called
Body schema	The fields the agent expects, shown as chips
Exposed to coordinator	Whether Flux may call it at all

Exposed to coordinator is the switch that matters. An agent can stay registered, with its credentials intact, while being invisible to the coordinator. Use it to park an agent you are not ready to let the model reach, rather than deleting and re-registering it later.

Editing an agent lets you rename it, update the description, or rotate header values. Last called is the field to check when cleaning up: an agent that has never been called either has a description the coordinator cannot route on, or is not needed.

Calling an agent from a chat

Two ways:

Let Flux choose. With the agent exposed and described well, Flux calls it when a request matches.

Call it directly. Type @ and the agent name:

@sop-agent what is sop?

Either way the call appears in the transcript as a tool card with its status, arguments, and an expandable result, and Flux writes the final answer from what came back. The Monitor link on the card opens that execution's record. See Monitor.

Agents, MCP servers, or skills?

Use	When
Agent	You own an HTTP endpoint or an A2A agent that already does the job
MCP server	The capability is exposed as MCP tools, often several related ones
Skill	The capability is a procedure over tools you already have, not a new endpoint

The distinction is where the work happens. An agent runs somewhere else and returns a result. A skill is instructions Flux follows itself.