A field guide · checked October 14, 2026
Agentic project management
Project work where AI agents read and write the project state directly, take multi-step actions on the team's behalf, and hand control back to people at the decision points the team has defined. Not a chatbot in a sidebar. A participant in the work, with limits.
#Definition
Three words carry the whole definition.
Most tools that call themselves AI-powered put a model beside the project: you ask it a question about the board and it answers. That is useful, and it is not agentic. An agentic tool is defined by what it is allowed to do to the project, and the definition turns on three words.
Directly
The agent reads and writes the real project state: the tasks, the people, the deadlines, the inbox. Not a copy pasted into a chat window.
Multi-step
It plans and executes a sequence (read the backlog, check capacity, draft the sprint, assign, notify) and carries the context across the steps.
Hands control back
At decision points the team defined, it stops and waits for a person. An agent without limits is a liability; an agent without autonomy is a script.
Researchers at Monash University describing the “agentic project manager” reach the same shape from the other direction: a multi-agent system that works like a junior or intern project manager alongside the team, with several working modes of varying autonomy chosen per task, and the human project manager moving toward the role of strategic leader and coach. The autonomy is a dial, chosen per kind of work. That dial is the product.
#Not the same thing
Assistant, automation, agent.
| Shape | How it works | Where it breaks |
|---|---|---|
| AI assistant | You ask, it answers. Summaries, drafts, questions about the board. | Nothing happens unless a person is at the keyboard. The coordination work is still yours. |
| Automation | If this, then that. A trigger fires a fixed action. | It cannot judge. The rule runs on the exception the same as the norm, and nobody wrote the rule for the case that matters. |
| Agent | Watches the project, decides when something needs attention, acts in steps, within limits a person set. | Without scopes, approvals and a ledger it is a fast new hire with every key in the building. The limits are the feature. |
#The autonomy ladder
How much the agent may do is a setting, per kind of work.
The mistake teams make is treating autonomy as one switch. It is a ladder, and different work sits on different rungs. Triage can run free; an email to a client should wait for a person. These are the four rungs Kavanah ships, from the bottom up.
Level 1
Read only
The agent answers questions, summarises and drafts, but writes nothing to the workspace.
Use it for: Day one. Enough to replace status meetings and the Friday report.
Level 2
Safe mode
The agent can create and update work, but sends, payments and deletes are stripped from every run.
Use it for: The first weeks of an AI employee. Most teams keep triage and planning here permanently.
Level 3
Approve
Whole categories of action (external email, finance, deletion) park for a human to approve before they run.
Use it for: Client communication, invoicing, anything that leaves the building.
Level 4
Full, within policy
The agent acts on its own inside the scopes, budgets and guardrails you set, and every action is logged and reversible.
Use it for: Internal coordination at volume: standups, nudges, forecasts, scheduled reports.
#What it looks like
Brief, work, ship. The agent carries the middle.
Brief
Describe the outcome in plain language, or forward the email. The agent turns it into a sequenced plan with owners, estimates and a charter it will honour.
Work
New requests get triaged into the right project. Tasks are assigned by skill and capacity. Blockers get chased. PR drafts appear on the tickets that are ready.
Ship
Standups, client updates and the weekly report write themselves from what actually happened, with sources, and land in the inbox before the meeting.
The agents are hired like teammates. An AI employee has a name, a role, a set of capability scopes, a reporting line and a schedule. You can assign it a task the way you would assign a person, @mention it in a thread, or put it on a Monday-morning run. It reports to a human, or to another AI employee who reports to a human.
#The controls
Lawful by design, or it does not ship.
Gartner expects more than forty percent of agentic AI projects to be cancelled by the end of 2027, and names unclear value and inadequate risk controls among the reasons. The agents that survive procurement are the ones whose every action traces to a policy a person wrote. These are the six controls every AI employee in Kavanah runs inside.
Capability scopes
Each AI employee carries a job description it cannot exceed. A CFO agent can read the books and never touch the send button.
Safe mode
One switch removes every external send, payment and delete from a run. The agent still plans, triages and drafts.
Action policies
Allow, budget-cap, park-for-approval or block a category of action per employee. An approval is a gate, not a suggestion: the agent stops and waits.
Guardrails
Block one specific tool for one specific employee, or require a condition before it runs. The rule survives model upgrades and prompt changes.
The ledger and undo
Every action lands in an audit ledger with who, what and why. The ones that can be reversed carry an undo, so a wrong move is a keystroke, not an incident.
A charter with a Negation
The KVN method asks every workspace to state what it refuses to do. The agent reads that before it creates work, and declines what the Negation forbids.
#How the tools compare
Ask every vendor the same five questions.
Does the agent act on project state directly? Can it run on a schedule without a prompt? Can an approval sit in front of a category of action? Is each agent scoped? Is there an audit trail with a way back? What follows is what each product's own page, or an independent 2026 review, said on October 14, 2026. “Not described” means the page did not say, not that the product cannot.
| Product | Acts on project state | Runs on its own | Approval gates | Scopes | Audit and undo | Pricing |
|---|---|---|---|---|---|---|
| Kavanah | Yes: plans, triages, assigns, drafts and reports across tasks, deals, mail and chat | Yes: scheduled runs, task-assigned runs, triggers on events | Yes: per-employee action policies (allow, budget, approve, block) and per-tool guardrails | Per AI employee, plus per-project and per-task access like a human member | Every action in a ledger; reversible actions carry an undo | From $3 per seat, AI usage metered in credits |
| ClickUp Super Agents | Yes: agents can be @mentioned, assigned tasks and run standups and triage | Yes: recurring schedules and an autopilot mode | Yes: reasoning logs and approval gates for high-stakes decisions | Configurable skills per agent | Reasoning logs | Brain add-on on top of paid plans |
| monday.com agents | Monitoring-led: progress, risk and capacity agents that alert and propose | Yes: briefers and nudgers run on their own | Not described; agents work inside the existing permission model | Per-agent read, write and act permissions | Not described on the agents page | AI on Standard plans and above |
| Notion custom agents | Yes, for knowledge-driven work; up to 20 minutes of autonomous execution per run | Yes: triggers on schedules or events | Not described | Workspace-wide per run | Not described | Agent credits beyond the base subscription; custom agents on Business and above |
| Asana AI Studio | Yes: creates tasks, reassigns work, posts updates | Not described | Not described; enterprise access controls and audit logging | Plain-language instructions per agent | Audit logging | Business tier and above; in beta |
#The first thirty days
Start narrow, read the ledger, widen.
- Week one
Connect the tools you already use. Hire one AI employee with a narrow scope: triage and the weekly report. Keep it in safe mode.
- Week two
Read its action ledger every morning. Where it was right, leave it. Where it was wrong, write the rule down as a Negation in the charter instead of correcting it by hand.
- Week three
Put the standup and the nudges on a schedule. Add an approval policy for anything that leaves the building, then let it draft client updates for you to approve.
- Week four
Hire the second employee for the next job (planning, or the inbox). Give the first one a reporting line to a human. Measure how many status meetings you cancelled.
#Questions
What people ask before they try it.
- What is agentic project management?
- Agentic project management is project work where AI agents act on the project itself rather than answering questions about it. The agent reads the current state of tasks, people, deadlines and messages, takes multi-step actions on the team's behalf (triaging a backlog, drafting a sprint, chasing a blocker, writing the Friday report) and hands control back to a person at the decision points the team has defined. The distinction from an AI assistant is that the agent is a participant in the work, not a feature you prompt.
- How is it different from the AI features in Asana, monday.com or ClickUp?
- A classic AI feature is reactive: you ask, it answers, you act. An agentic tool monitors the project continuously, decides when something needs attention, and acts within limits a human set. Several established tools now ship agents of their own, so the useful question is no longer whether a product has AI but how its agents are scoped, whether an approval can sit in front of a category of action, and whether every action lands in an audit trail you can reverse.
- What can an AI project manager actually do on its own?
- In Kavanah an AI employee can plan a sprint from the backlog and capacity, triage new requests into the right project, assign work by skills and availability, draft pull requests and client updates, run a daily standup summary, chase overdue items, keep a revenue forecast current and write the weekly report. Which of those it may do without asking is a setting, not a property of the model: each AI employee carries capability scopes, an autonomy level and action policies.
- How do I keep an agent from doing something we decided not to do?
- Three layers. A safe mode strips sends, payments and deletes from a run entirely. Action policies let you allow, budget-cap, park-for-approval or block a whole category of action per AI employee, and a guardrail can block one specific tool for one specific employee. And a workspace charter with a Negation (the KVN method) tells every agent what the organisation refuses to do, which the agent must honour before it creates work. Everything an agent does is written to a ledger, and the actions that can be reversed carry an undo.
- Does agentic project management replace the project manager?
- No. Researchers describing the agentic project manager compare it to a junior or intern project manager working alongside the team, with the human moving toward a strategic and coaching role. What goes away is the status-chasing, the manual triage and the report assembly. What remains is judgment: priorities, trade-offs and the hard conversations. The manager's job gets denser, not smaller.
- What does it cost to run?
- Kavanah is priced per seat from $3 a month, with AI usage metered in credits on top, so a team pays for the agent work it actually runs rather than a flat AI add-on. A typical chat reply costs around 50 credits; an autonomous scheduled run costs more because it reads more of the workspace. The pricing page has a calculator.
- Does it work with the tools we already use?
- Yes. Kavanah connects to GitHub, Jira, Asana, ClickUp, Monday, Notion, Google Workspace, Microsoft 365, Slack and Telegram, among others, and the agent only gains a provider's tools once that provider is connected. You can keep Jira as the system of record for engineering and let the agent work across it.
- How do I start?
- Create a free workspace, connect the tools your team already uses, and hire one AI employee with a narrow scope, for example triage and the weekly report. Keep it in safe mode for the first week, read its action ledger, then widen the scope. Most teams reach a point within a month where the standup, the triage and the status report run themselves.
Sources
- Toward Agentic Software Project Management: A Vision and Roadmap — arXiv (ICSE 2026 AGENT workshop preprint)
- Agentic Project Management: How AI Agents Change Where Coordination Lives — Quire
- Best AI Agents for Project Management in 2026 — Fastio
- AI Agents for Project Management: Build your own — monday.com
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — Gartner
The busywork runs itself.
Free to start. Hire one AI employee, keep it in safe mode, and read what it did. Related reading: the KVN method, the product and the blog.