You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
REST API: call your agents from your own systems
Major New Feature. Only Company Admins.
Your own systems can now call an agent over HTTP and get back a JSON answer in a shape you define.
Switch it on under the agent's Configuration → Channels → REST API, and choose which kinds of caller it accepts. API keys and OAuth clients can call only Company agents, so release a personal agent to your company first.
Publish a contract: the input parameters the agent takes and the JSON Schema its answer must satisfy. A published version is frozen; to change it, publish the next version and deprecate the old one with a sunset date. Callers pinned to it get
DeprecationandSunsetheaders until then.Test it with a real call on the agent's Test API tab, then hand your integration the agent ID, the contract version, and an API key or OAuth client.
Coming soon: MCP support, so AI assistants and IDEs can call your agents as tools, and an OpenAPI spec generated from each agent's contract.
Learn more: Setting up an agent's REST API and the developer guide, Agent REST API.
The Test API tab: pick a contract version, fill in the inputs, click Run and check the agent's JSON answer.
Skills: a new Italy round-trip sample itinerary (TAIDEMO-ITALY)
New Feature.
TAIDEMO-ITALY is a 14-day round trip from Zürich through Rome, Florence, near Bologna and Venice: flights both ways, a rental car for the whole loop, hotels, guided tours and lunches. It comes from a real TripBuilder itinerary that is still being planned — nothing is booked and no traveler details are on file — so you can see how the checks read a trip at the planning stage. Try it with any agent equipped with the Sample Itineraries skill.
Connectors: reopen an active connector to fix or extend it
New Feature. Only Company Admins.
Click Revise with AI on a connector built in the Connector Builder: the assistant starts from the live code, reproduces where it stands, and proposes only the change.
Before you activate, the review shows what the revision changes (operations, hosts, code, credential fields) and, when an operation is removed or changed, which agents and skills it affects. Agents keep using the active version until then.
A Versions section on the connector's page lists every version with its code.
Learn more in the Connector Builder guide.
A revision starts from the active version: the assistant runs it unchanged first, and agents keep using it until you activate.
The Versions section: which version is active and what each one was based on.
Harness: bulk lookups now work with Connector Builder connectors
Improvement.
When an agent looks something up for many items in parallel, each item can now read from a connector built in the Connector Builder too.
Agents: Traveller Locator finds who is where by night, tour route and travel day
Improvement.
The Traveller Locator used to find travellers by the dates of the services they booked, so a free day in a city or a tour that only passes through could be missed.
It now works from where each booking actually is on the date you name — where the travellers sleep that night, where they travel that day, and where a tour or day excursion takes them during the day. Results come in three groups: in the area, visits the area during the day, and possibly in the area when the itinerary doesn't say where they sleep. You can also search around a place with a radius, such as "within 80 km of Mount Etna". The detail pass now confirms each booking live and flags any cancelled or withdrawn since the data was last refreshed.
Harness: long lookups no longer report skipped items as failed
Fix.
When a turn hits its tool-call budget, the answer now names the lookups that were not attempted instead of calling them failed.
Agents: saving an agent no longer widens the tools it may use
Fix.
A Workbench save dropped an agent's tool restrictions, so copies of built-in agents gained tools they were built without.
Agents no longer offer to run code they cannot run
Fix.
ThunderAI agents no longer mention shell or Python tools they do not have, so they stop proposing code runs that cannot happen.