Yes, this is Kavanah's blog, and yes, Kavanah is number one. Read that ranking with whatever salt you like. The other nine are real products, described from their own pages, which we fetched and read on October 22, 2026. We ranked them on the four questions from our agentic project management page: does the agent work on the project state directly, are its permissions scoped, is there an audit trail with a way back, and can a person intercept an action before it commits. The thing most demos skip is that fourth one. A vendor page that does not mention an approval step is not proof the product lacks one, so where a page is silent we say the page is silent, not that the feature is missing.
1. 1. Kavanah: Agents Built Around the Brakes
Our own product, so here is the claim in full and you can test it. An AI employee in Kavanah has capability scopes, an autonomy level from read-only to full, action policies that allow, budget-cap, park-for-approval or block a whole category of action, and guardrails that block one specific tool for one specific employee. Per-project and per-task access applies to agents exactly as it does to people, and every action lands in a ledger where the reversible ones carry an undo. That is a yes on all four questions. It suits teams that want agents running triage, planning, reports and client drafts on a schedule, with external email waiting for a person. The honest limitation: Kavanah is younger than everything else on this list, and its integration catalogue is deep where agentic work needs it and thinner at the edges.
2. 2. ClickUp Super Agents: The Closest Big-Tool Match
ClickUp launched Super Agents on December 28, 2025, and its launch post describes them as AI teammates modelled as real users: you assign them tasks, @mention them in docs, tasks and chats, DM them, or put them on schedules and triggers, such as triaging new requests as they arrive. They inherit the same user-level permissions as team members, every action is recorded in real-time audit logs, and for high-impact actions they request human approval, with reviewing an email draft before it is sent as the example. That is the clearest four-for-four story of any large vendor we read. It suits teams already living in ClickUp. The limitation is that the launch post does not list plan tiers or prices, so budget for a conversation with sales before you budget for the agents.
3. 3. Notion Custom Agents: Best Paper Trail
Notion's custom agents run on a trigger or a schedule without anyone present, with triggers from Slack reactions and mentions, new mail from specific senders and calendar changes. Each agent gets its own permissions, like a teammate, down to page level. Notion is the most explicit of the big tools about the after-the-fact side: every agent run is logged with what triggered it, what it did and why, any change can be undone through version history, and admins can stop every run for an agent. It suits knowledge-heavy teams whose project lives in docs. The limitation: the page describes no general approval step before an agent acts, and custom agents sit on Business and Enterprise plans and, from May 4, 2026, run on credits sold as an add-on at ten dollars per thousand a month.
4. 4. monday.com Agents: The Biggest Gallery
monday.com lists a long shelf of ready-made project agents, among them a Project Intake Agent, a Status Reporter, a Nudger, a Dependency and Risk Mapper and a Resource Rebalancing Agent, plus a no-code builder for your own. Its page says you control exactly what each agent can read, write and act on, within monday's existing permission model, and that agents are on all paid plans from nine dollars a seat a month, with each action running on AI credits. It suits teams that want a working agent this afternoon without designing one. The limitation: the page does not mention audit logs or undo, and its Approvals Agent routes human sign-offs on requests rather than gating the agents' own actions.
5. 5. Jira with Atlassian Rovo: Approval by Default
Atlassian describes Rovo agents as configurable AI teammates any team member can call on or create, usable in chat, in automation rules, and while editing in Jira and Confluence. The detail that earns the placing is in the wording: agents perform skills like organising, creating or editing Jira work items with your permission, and act outside chat with your confirmation. For a software team already on Jira, that is a low-risk way to start. The limitation: the support page we read does not describe a ledger with undo or how per-project scoping works for an agent, and Atlassian's Rovo page did not settle which plans include it, so check the licensing page rather than assume.
6. 6. Asana AI Studio and AI Teammates: Workflow-First
Asana has two layers. AI Studio is a no-code builder for automations triggered by a new task, a form submission or any project event, with actions from routing work to the right owner or approval stage through to creating and reporting; you control exactly what data the AI can reference, and a Basic tier is included on all paid plans with rate limits. AI Teammates are thirty-plus prebuilt agents, Project Manager and Insights Analyst among them, that see only the projects your existing permissions allow. It suits operations teams who think in workflows. The limitation: neither page describes an audit log of agent actions or a formal approval step for Teammates, despite a heading promising you are always in control.
7. 7. Linear: Agents That Never Take the Ticket
Linear treats agents as full members of the workspace that can be assigned issues, added to projects and @mentioned, and lists a roster of integrated ones: its own Linear Agent for triage, plus coding agents such as Cursor, OpenAI Codex, Devin and Factory, and Sentry for root-cause analysis. The design choice we like is accountability: when an issue is delegated, the human stays the primary assignee and the agent is added as a contributor, and Linear says agents act on your behalf but never in the dark, with each change and its reasoning visible. It suits engineering teams who want coding agents fed from the tracker. The limitation: the page does not describe permission settings, approval workflows, audit logging or pricing for agents.
8. 8. Wrike AI Agents: Enterprise Chains
Wrike made its agents generally available on February 5, 2026, with three prebuilt ones, Risk Status Reporter, Triaging and Intake, an Agent Builder, Multi-Actions so one agent can run several steps, and Agent Chaining to link agents in sequence. Only space admins can create and deploy them, builders can simulate and tune an agent in a sandbox first, and users can see the reasoning behind an agent's actions. It suits larger organisations that want central control over who ships an agent. The limitation: agents are on the Business, Pinnacle and Apex plans only, and the announcement does not describe a human approval step or an audit log.
9. 9. Quire: Bring Your Own Agent
Quire's angle is different: rather than a built-in agent zoo, it offers an MCP connection so the agent you already use can read and update your real Quire tasks. Its own playbook, which coined the four-question test we use, says permissions can be scoped per project and every action is logged and reversible, and counts Quire among the few tools that pass. It suits small teams that already run an assistant like Claude and want it working on a simple, clean tracker. The limitation: those are Quire's claims about Quire, written by its marketing team, and the post does not explain how a person intercepts a planned action before it commits.
10. 10. Dart: An Agent in Every Workspace
Dart ships Dart AI, a built-in project management agent in every workspace that creates tasks, tracks progress and manages deadlines, with adjustable defaults. You can write custom agents that start working the moment work is assigned to them, for jobs like estimating sizes or breaking a deliverable into subtasks, and route work to external agents such as Cursor and GitHub Copilot. It suits small product teams who want agents as assignees from day one. The limitation: its help article does not describe permissions, review or approval, or audit logging for agents.
11. Near Miss: Motion
Motion is the tool people most often expect on a list like this, and its AI Project Manager does act on its own: automatically updating statuses, assigning next steps and flagging delays, with AI models covered by the subscription. It missed the ten on the test, not on quality. Its page describes a quick review of AI-built projects but no approval gate on automatic status changes or reassignment, and no role-based permissions for the AI. Teams whose real problem is a personal calendar rather than team coordination may still like it best.
12. How to Use This List
Do not pick from the ranking; pick from the gaps. Every entry above has one, ours included. Take the four questions to a fifteen-minute call with the two vendors closest to your current stack and ask them to show, not say, three things: the setting that stops an agent from sending an email until a person approves it, the screen where you read what the agent did yesterday, and the button that reverses one of those actions. If a vendor can show all three, you have found an agentic tool. If the answer is a demo of how clever the model is, you have found an assistant with a scheduler.
Vendor pages change fast in this market, and several here launched their agents inside the last year. We read every page on October 22, 2026, and we will re-check them rather than let this list quietly age. The longer version of the four-question test, with the autonomy ladder and a side-by-side table, is on our agentic project management page.
