This week, project agents became easier to direct and easier to understand. You can give them one project goal, shared instructions and reviewed memory, then choose exactly which events start them.
Work feels more immediate, too. Boards, sprints and open tasks now update as teammates and agents make changes. We also shipped improvements to systems, webhooks, AI models, billing and the app’s supported languages.
Give project agents a clearer brief
The rebuilt project Agents tab puts each agent’s briefing in one place. Set a goal and instructions shared by every agent on the project, then choose each agent’s triggers, approval level and individual instructions. The tab also shows its tools, skills, connected apps, model, recent runs and token use. Goals previously set on individual project agents were moved to their projects, and project instructions now reach agents started by their own event triggers.
You can review project memory, correct or archive facts, and pin facts that must always be included. Agents receive the project’s ten most important memories, plus pinned ones. Corrections and preferences you state are retained long-term, and facts recorded on project tasks can become project memory. A duplicate memory topic now produces a clear error instead of silently competing with an existing fact.
Choose whether assignment, an @mention, a new or changed task, a completed task, or a new comment starts each project agent. A table shows which events start each agent, and an agent with no selected triggers stays inactive on that project.
Narrow the task-updated trigger to fields such as title, description or scheduled date. Related edits are bundled before an agent starts, and it receives what changed, who changed it and the final task state.
Test an agent against a project task and event without changing or posting anything. Recent runs show what started them and how they ended; an orange spinner shows where an agent is working.
Assigning or mentioning an agent already on a project now respects those triggers; active agents keep both triggers selected unless you change them. Task-completed triggers also fire when a task is simply marked done. Approval labels now describe when approval is requested, and project-level token budgets, which were not enforced, are gone. Agent budgets remain on the agent. Adding or managing project agents now requires Advanced AI, rather than a separate project-owner permission.
Follow task and project work as it happens
Project boards, lists, calendars, Gantt views and Sprints now refresh as work changes, including changes to tasks, columns, sprints and comments. An open task updates too, including its comments, dates, assignee and subtasks, while preserving what you are typing. You will see a Disconnected label if the live connection drops; private-project updates remain limited to people with access.
The new Milestones page brings milestones from all projects into one view, with dates, progress, overdue status, filters and expandable task lists. You can also switch an existing project between public and private after previewing who gains or loses access, including task assignees. Making it public requires confirmation; existing project content stays put, and people previously granted access retain it if you later make the project private again. MCP assistants can preview and request the same change, and can identify the project owner.
AI-made task changes now follow the same completion, notification, activity, milestone and recurrence rules as changes made in the app. Activity badges distinguish work done through an AI assistant or agent, and task deletion appears consistently in the audit log. We also corrected task links, cross-organization navigation and moves, template-created subtasks and dependencies, linked-task board positions, blocker notifications, comment-only permissions, task-panel switching and sprint filters.
Dates and everyday editing received attention too. Due dates and time logs use your timezone correctly; recurring tasks retain their local time, and schedules understand phrases such as “monthly” and “every 2 weeks.” The emoji picker, numbered and bulleted lists, narrow board cards, dropdown dismissal and linked documents now behave as expected. Google Tasks sync now carries scheduled dates separately from due dates, retries missed edits and reflects OKiDO task changes more reliably.
Make runs, systems and connected tools more dependable
In the system builder, a new Merge step lets a run continue when either decision outcome arrives. The builder warns about a Join that could wait forever and prevents publishing that configuration; valid parallel Joins still wait for every branch. We also fixed parallel execution, loops, gates, approvals, task-created steps and progression after a skipped or completed step. Interrupted steps and runs recover more reliably, and completing a cancelled run’s last step no longer changes its status to completed.
AI assistants can now start SOP runs with their variables and full step details, follow step order, read run variables and find templates in department processes. They can set or clear task dates and assignees, see recent comments first, and receive usable links to records. MCP assistants can also file and check support tickets without creating duplicate submissions. They now explain task closures, retain durable project facts and confirm before deleting or contacting others; agent requests, approvals and failures return clearer results.
Signed webhooks now deliver subscribed events, including changes to requests for approval, and show recent attempts, responses and a resend option. Failed deliveries retry, while private-project events require an explicit opt-in. If you used an existing webhook, reveal its new signing secret in Settings and update your receiver. Run cancellations and tasks deleted with a project now produce the expected events. Older, unused webhook settings have been retired.
Schedules now run at the intended local time across daylight saving changes, reject unclear phrases and respect concurrent-run limits. Run steps record who completed them, and run status controls appear for the people allowed to change them. Old code-based SOP step automations have been retired: assign an automated step to an AI agent instead. Decision trees, escalation alerts, reminders, timers, sprint reports, CSV exports, recordings, imports and the Capability Factory also received fixes for incorrect results, missed work or unclear errors.
Choose AI models and access with more confidence
OpenAI Sol 6.1 is available to Pro and Business organizations. Anthropic Fable 5.1 and Opus 5.5 are available on those plans, while Sonnet 5.5 starts on Starter. Opus 5.5 uses 20% fewer allowance tokens than Opus 5 for the same work; we also corrected Sonnet 5’s allowance rate.
Organization AI settings now determine the models used for chat, task agents and MCP-started agents, unless an agent has its own model selected. New defaults include Luna 5.6 for chat and Sol 6.1 for agents and MCP where the plan allows; Free and Starter fall back to an included model. Agent pages show the organization default and only offer included models. The unused Coding default is gone: Capability Factory stages use their own model settings. Existing model choices remain in place where available, and unavailable choices fall back to an included model.
Owners and admins can grant Advanced AI per member. It opens agent and advanced AI management areas; members without it retain Chat, Skills and Memory, and can still talk to agents and assign them tasks. Everyone can import skills and edit their own unused skills. We also fixed agent replies, approval recovery, run status, budgets, scheduled jobs, permission removal and capability builds, including cases where a failed run appeared to keep working.
Find your way around, in your language
The app is now translated across its main workflows into Spanish, French, German, Dutch and Portuguese. Counts, statuses and priorities read naturally, Spanish screens no longer mix languages, and German and Dutch use consistent informal address.
Labels can now be grouped into categories, with a searchable multi-select picker for tasks, documents, templates and recordings. MCP tools can read categories and create labels within them. Dashboards are personal to each member, with a starter layout for new members. The organization switcher makes choosing one or several organizations clearer, and links, invitations and settings work more consistently when you manage multiple organizations.
Billing and usage are clearer too. Admins can manage subscriptions; paid-plan owners and admins can buy storage and AI token packs. Existing uploads now count toward storage, while their links continue to work. Plan limits apply to more creation, import, scheduling and AI actions, and additional AI features count toward the monthly allowance. Public run sharing now needs a paid plan, although existing links keep working. We also corrected checkout, add-on cancellation, overage charges, duplicate charges and billing errors.
Finally, API tokens now respect their granted resource scopes; new tokens require an eligible paid plan, and password resets or signing out everywhere revoke personal tokens. Existing integrations may need additional scopes. Search returns the first 1,000 matches, so refine broad searches, and the old documents address now opens the Playbook. Across the app, we fixed broken links, upload and OAuth flows, accessibility labels, validation messages and interruptions that could leave work waiting without an answer.
To try the biggest changes, open a project’s Agents tab and review its shared briefing, triggers and memory, then run a test against a task. Keep a board and task open while your team works to see the live updates, or visit Milestones for a cross-project view of what’s next.
