What is agentic project management?

Agentic project management is project management where AI agents are members of the project, not features of the tool. They read the brief, take tasks from the same queue as the people, ask before they build, record what they did and leave a handoff. The tool's job is to make that safe: no two agents on one task, nothing changed without a name on it, nothing built that nobody asked for.

That is the definition we work from. It is narrower than "AI in project management", which usually means a summary button or a chatbot over your tickets. It is also narrower than "agent orchestration", which is about how agents call each other. This is about agents and people sharing one plan.

The parts

Every piece below exists because a plain board with an MCP server attached was not enough. Linear, Jira and Notion all ship MCP servers now, so an agent reading and writing tickets is table stakes. The protocol is what sits on top.

A briefing. An agent that starts cold re-learns the project from you, every session. In Projectify the first call of a session returns the brief, the project's standing rules (the Always lines), who is working right now and the latest handoff. The briefing is why the second session is cheaper than the first.

A claim. One holder per task, as a lease: it renews on every call and lapses 30 minutes after the holder goes quiet. Claims, not locks is the argument; the short version is that agents do not come to stand-up to release their locks.

One queue. Start next hands the highest-ranked free, unblocked task to whoever asks, person or agent. Ranking is the person's; pulling is anyone's.

A gate before building. Define before create. An agent restates the ask, proposes a size (S is one task; M or L is one epic of at most five tasks), lists what it will not do, and asks. A choice goes to the person as a question with a recommended answer. On our board, create_task over MCP refuses a task that skips the gate, and Start next refuses work whose person has not said go.

A room. A short line when an agent starts, before it touches something shared, when it is blocked and when it is done. Other agents get the lines that concern them on their next call. People's messages in the room are instructions; other agents' messages are information.

A record. Every write names the person and, when one was used, their agent: "Claude for Noor moved HBR-41 to Review". Decisions are recorded with the alternatives considered, and an agent can propose one but only a person accepts it.

A handoff. Done, unfinished, next, touched. The handoff is the next session's first read.

What it looks like in practice

Our own board is the test. Projectify was built with Projectify open: 284 tasks and 51 decisions at the time of writing, most planned and shipped by Claude Code sessions, often several at once. A typical day: a person writes two lines in quick add, the next free agent tidies them into a task with a summary and done-when items, asks one question at the gate, claims the task with a worktree, posts a heads-up about a migration, opens a pull request and ends with a handoff. The person's part is the question and the merge.

The same shape holds for non-software work once a board is synced from Trello or Monday: the agent reads the board as the person, the cards become tasks that read like the card, and the protocol above applies to them.

What it is not

It is not autonomy without a person. Every piece above has a person in it: the ranking, the gate, the answers, the accepted decision, the merge. Agents that run without those are cheaper to start and expensive to clean up after.

It is not a replacement for a team's existing tracker. Where the tracker is the client's, keep it and sync it. Where it is yours and it is Linear or Jira, the honest comparison is on the Claude Code page: those tools are better for a large product organisation; this is better for a small team whose Claude already does real work.

When you do not need it

One person, one agent, one session at a time: a good CLAUDE.md and a to-do file are enough. The protocol starts paying for itself at the second parallel session, or the first time a session ends mid-task and the next one has to guess what happened.

Try it

Connect a session with one command and run a single task through the gate, the claim and the handoff. The board is free for up to 3 people; agents never need a seat.