AI agents that work inside your app
An AI step without tools answers once and hands back text. Give it tools and it keeps going, turn after turn. It can call your integrations or run another workflow, build and edit things in your workspace, or open your site in a browser and look at it, until the job is done.
The block is on every plan, including Free. A paid plan lets each run go longer before it hands off to the next.
Try it
An agent with tools, at work
Prompt
'Check the shop pages and list anything broken.'
Tools
Scope
Max turns
Model calls, not tool calls
Ready
Press Run to watch the agent work through the task, one model call at a time.
Change its tools, scope or turn limit first and the run changes with them.
How it works
- 01
You choose its tools
An agent gets only the tools listed on its step, down to which Tessryx operations it may use. Firing a schedule and changing members or billing aren't available to it.
- 02
It's limited to the folders it's given
An agent that can build needs a scope, a list of folders it may touch. It can't widen the list, and whatever it creates is held to it.
- 03
It can use a key without seeing it
If a workflow holds your image service's key, the agent can call that workflow without ever seeing the key.
- 04
Jobs longer than one run
Near the time limit the agent stops at the end of a turn, and the next run in the chain picks up the same conversation.
Questions
Which models can it use?
Anthropic's models, with your own API key. It can also call a model server you run yourself, if that server speaks Anthropic's API.
Can it break things outside its folders?
No. Its tools only reach the folders in its scope, and every save is checked against that scope first.
Under the hood
The technical details
- tools
- Any mix of: one of your integrations, another workflow, Tessryx operations (build, edit, publish, run and preview in the workspace), and page snapshots.
- max_turns
- Model calls, not tool calls: 2 to 50, required with tools. Running out of turns while still asking for tools fails the run.
- limits
- Up to 32 tools per block. Tool calls within a turn run one after another, and prompt caching is applied for you.
- tessryx
- Each ticked operation becomes a tool. The agent acts with a credential minted for the run and is refused before it starts if it has no scope.
- recursion
- A run or preview an agent starts counts toward the same limit as the run it is inside, so it cannot loop back in through the front door.
{
"scope": { "grants": [{ "prefix": "shop", "actions": ["read"] }] },
"secrets": ["anthropic-key"],
"steps": [
{
"name": "review",
"blocks": ["llm"],
"llm": {
"provider": "anthropic",
"model": "claude-sonnet-5-5",
"secret": "anthropic-key",
"max_tokens": 32000,
"max_turns": 12,
"prompt": { "language": "jsonata", "source": "'Check the shop pages and list anything broken.'" },
"tools": [
{ "kind": "tessryx", "tessryx": ["list_dynamic_endpoints", "preview_dynamic_endpoint"] },
{ "kind": "page_snapshot" }
]
}
}
]
}Related features
Scoped agents
An agent that can build is limited to the folders it was given. It can't add to that list, and whatever it creates or connects has to fit inside it.
Long-running agents
An agent can keep working across many runs. At the end of a turn it hands its conversation to the next run, and older turns get summarized to keep the conversation small.
Agents that look at your pages
An agent can open your site in a real browser and check its own work. It gets screenshots at desktop and phone sizes, plus a list of any text that overlaps or runs off the edge.
AI steps in a workflow
Any step can ask a model to write or summarize something, and the answer flows into the rest of the workflow. It runs on your own model account, with no markup from Tessryx.