Changelog
Product and API updates, published twice monthly. Every release is tagged: filter by product area or search the full archive to find changes relevant to your work.
Entries describe changes at the time of release. Follow each entry’s current documentation links for today’s setup, API behavior and model availability.
Evals and Run Operations
What’s New
File-Backed Task Evals
Task evals can replay reference runs that used direct image or PDF inputs. Stored processed values are reused, and direct model-facing files are restored from Project storage for the candidate run. See Evals.
Detailed Agent Run Operations
Connections, Approvals, and Xero
What’s New
Personal and Shared Connections
Supported integrations can now hold both a user-owned connection and a shared Project connection. Each Task or agent binding selects the invoking user’s credentials or the Project default, and the integration catalog reports the scopes and policies each provider supports. See Connections.
Model Governance and Evaluations
What’s New
Model Availability Controls
Organization and Project administrators can exclude models by model ID, provider, or vendor. Organization rules flow into every Project, and safety checks prevent a policy from removing every active model. See Model governance.
Approvals, Chaos Mode, and Audit Integrity
What’s New
Approvals in the Dashboard
Human-in-the-loop approvals now have a home in the dashboard. A new Approvals tab surfaces every pending request across your agents, so a reviewer can see what a tool is about to do and approve or reject it. An approved API run resumes when the client calls the resume endpoint. The record captures who decided what, with recognized secrets redacted from its argument preview. See Approvals.
Slack, Mistral, and File Attachments
What’s New
Slack Integration
Connect Slack as a Connection and let your agents work inside your workspace. Agents can list channels and post messages through your connected Slack app, and the dashboard shows each connection’s authentication status with clear prompts when a reconnect or scope upgrade is needed. See Connections.
Approvals, Runtime Files, and Gmail Triggers
What’s New
Human-in-the-Loop Approvals
Gate any agent tool call behind a human decision. Mark a tool as requiring approval and the run pauses when that tool is about to fire, so a person can approve or reject the action before it happens. This is built for tools with real-world side effects, like sending an email or updating a record. See Approvals.
Context Usage, Audit Log, and File Attachments
What’s New
Context Usage Breakdown
See exactly what fills an agent’s context window on every run. A new usage panel attributes input tokens across instructions, conversation history, tools, Skills, and integrations, and an expandable context chip highlights the breakdown at a glance. Usage is aggregated across the runs in a session. See observability for details.
Evals, Agent Sharing, and Memory Compaction
What’s New
Evals for Tasks and Agents
Test changes before you ship them with evals. Replay real past runs against a candidate revision and let a judge model score each result as pass or fail, so you catch regressions before they reach production. Eval pages are now shared across both Tasks and agents.
Skills, Connections, and Agent Versioning
What’s New
Skills
Give your agents reusable, focused abilities with Skills. Browse ready-made Skills from the marketplace or build your own with instructions and reference files, all managed through a new create-and-detail interface. Skills are versioned, so you can iterate safely and pin the exact revision an agent uses. Skills can also declare the tools they depend on, and those dependencies are provisioned onto your agent automatically when the Skill is attached.
Integrations and Agent Experience
What’s New
Integrations
Integrations are native, platform-managed connections to external services, starting with Google Workspace: Sheets, Docs, Slides, Calendar, and Gmail. Your agents can read and write these services directly, without you standing up an MCP server. Integrations join tasks, Skills, and MCP servers as a way to give agents tools. See Connections.
Skills and Knowledge Bases
What’s New
Skills
Skills are reusable, versioned abilities your agents can call on to complete focused work. Browse a marketplace of ready-made Skills, or create your own project-scoped Skills with their own instructions and reference files. Skills are added to an agent alongside its tasks and tools. See Skills.
Agent Triggers and Execution Modes
What’s New
Triggers for Agents
Triggers can now start an agent, not just a task. Fire an agent from an incoming webhook, or run it automatically on a recurring schedule with cron and time zone support. Webhook-triggered agent runs are accepted immediately and processed in the background, so long-running work never blocks the caller. See Triggers and runs.
Agents and MCP Connections
What’s New
Introducing Agents
Agents are AI orchestrators that reason over natural language and call your existing tasks to get work done. An agent can select and run multiple tasks in sequence, hold a multi-turn conversation, and stream its responses as it works. Build one in the dashboard and give it the tasks it needs. See Agents.
Triggers and Automation
What’s New
Triggers and Webhooks
Tasks can now run automatically in response to external events. Create a webhook endpoint that runs a task whenever an outside service calls it, map the incoming payload to your task inputs, and secure it with HMAC or bearer token authentication. Connect GitHub, Stripe, Slack, or any service that can send an HTTP request. See Triggers and Runs.
Visual Inputs for Tasks
What’s New
Image and PDF Inputs
Tasks can now accept images or PDF documents as visual input for the model to read and reason over. Set how a task handles files with the new file input mode, choosing between images only, PDFs only, or both. This makes it straightforward to build tasks that check text against a screenshot, extract fields from a scanned document, or validate visual content. See Tasks.
Image Generation and Agent Protocol
What’s New
OpenAI Image Generation
Create images directly in your tasks using OpenAI’s GPT Image model. Configure size, quality, and output format. The AI generates images from your prompts with support for various aspect ratios and quality levels.
A2A Protocol Support
Email Forwarding and JSON Schemas
What’s New
Email Task Forwarding
Forward task results directly to email addresses. Configure one or more recipient email addresses and the task output will be sent automatically when the task completes. Uses a template editor with variable suggestions to customize the email content.
New Output Formats
What’s New
PDF Output
Tasks can now generate PDF documents directly. The AI creates HTML content that’s automatically converted to PDF files. Perfect for reports, invoices, and formatted documents.
CSV Output
Share Examples and Download Results
What’s New
Share Tasks with Example Runs
When sharing tasks publicly, you can now attach an example run to show exactly what the task produces. Viewers see the example inputs and outputs, and when they clone the task, those inputs are pre-filled automatically.
Task Cloning and Sharing Controls
What’s New
Task Cloning
Duplicate any task with a single click. Clone tasks across projects - works for logged-in users, registered users who need to sign in, and new signups. Perfect for creating variations or sharing templates.