What’s New in Agent.ai: Product Updates, Features & Improvements

There’s always something new to explore in Agent.ai.

And now you'll know just where to find it.

Every time we ship a feature, refine a workflow, or smooth out a rough edge, we’ll add it here. Some updates will be big. Others will just make your day that much easier. All of them are designed to give you smarter agents, help you discover better agents, and make sure you get more value from the agents you use.

Read on for what’s new.

🔐 Google Docs-Style Agent Sharing

Agent sharing just got an major upgrade.

We’ve rebuilt agent sharing from the ground up. The old “Sharing & Visibility” dropdown in the agent experience is gone. In its place you'll find a dedicated Share tab that works just like sharing works in Google Docs.

For Builders

You can now:

• Share agents by email by assigning one of three sets of permissions by role:
• Viewer - run only
• Commenter - run + view configuration
• Editor - run + view + edit configuration
• See exactly who has access, complete with real names and avatars
• Change roles or remove access instantly
• Choose Anyone with the link or Restricted access
• Submit agents to the public directory directly from the Share tab
• Control agent-to-agent invocation with a simple toggle
• Changes save instantly — no more “Publish Changes” for sharing settings

For Agent Users

Visiting a restricted agent now looks and feels more familiar. You’ll see a clean “You need access” page where you can request Viewer, Commenter, or Editor access — with an optional note.

Owners get an email, approve or deny from the Share tab, and once approved, the page updates automatically. No refresh. No guessing.

In Your Agent List

• Shared agents now appear in your agent list with clear role badges
• Clone is available to Editors and above
• Delete is reserved for owners
• Status now shows Restricted (instead of “Restricted Users”)

It’s cleaner, faster, and much harder to share something with the wrong permissions.

✉️ Custom Instructions

Too often working with AI means repeating yourself — instructions, background information, other kinds of context — over and over again. Now you can teach your agent your preferences just once. And move on.

There’s a new Email Agent tab in Settings where you can define how the agent should respond: tone, format, language, structure — whatever matters to you.

No instructions set? Nothing changes.

Less re-explaining. More consistency.

🔌 OpenClaw Actions

Workflows can now talk directly to OpenClaw.

OpenClaw actions are live in the Action Library, which means builders can connect workflows to a user’s personal OpenClaw instance without duct tape or manual config.

Three new actions:

Get Details
Drop it in first. It automatically resolves the user’s instance credentials — no IPs or tokens required.

Chat Completion
Send messages through the standard OpenAI-compatible API. Supports multi-turn conversations, session persistence, and streaming.

Tools Invoke
Call any tool available on the user’s OpenClaw Gateway directly from a workflow step.

You can find all three actions in the new OpenClaw category. 

🧠 Knowledge Agent Evaluations

Now you can turn regular conversations into real answers.

Knowledge agent builders can now review user conversations and curate few-shot examples directly from the builder.

There’s a new Evaluations tab under Knowledge & Eval that shows message pairs from real interactions.

Builders can now:

• Toggle messages on or off as examples
• Edit responses to create ideal answers
• Automatically inject selected examples into the agent’s system prompt at runtime

By the numbers:

• Up to 100 examples per agent
• 5,000-character budget for prompt injection
• Polls for new messages every 10 seconds
• Zero impact if no examples are configured

It’s agent training that's grounded in reality, not guesswork.

⚡ Event Triggers UX Overhaul

Triggers now live right where they belong.

You can browse, connect, and configure event triggers directly from the Edit Trigger sidebar. No more hopping into the Actions Library to wire things up.

What’s new:

• Third-party action vendors show their connection status upfront
• One-click Connect handles OAuth inline
• Required fields are validated before creation, with clear, human error messages
• Active triggers display in the builder as:
  Trigger Agent On: Email Sent (with a clear icon)

🤖 Agent Builder Assistant — Major Upgrade

The Agent Assistant now generates higher-quality agents, faster, and with far fewer surprises.

For builders:

• Actions are generated with guaranteed valid structure
• Supports complex workflows with 2x larger output capacity (32k tokens) using the latest Claude Opus 4.6
• Self-correcting generation—the assistant validates and refines before delivering
• Live progress updates during generation
• Generated actions save immediately and persist across refreshes
• Fully editable after applying

On reliability: We fixed several edge cases around action editing, template variable validation, and draft persistence. Added 71 automated tests covering the full generation-to-validation pipeline.

Smarter generation. Stronger guardrails. Better agents for you.