Every tool now claims to be "agentic". Most of them mean "we added a chat box". Here's the definition that actually matters for project work: an agent reads live state, performs multi-step work on it, and hands control back to a human at defined checkpoints. A chatbot answers questions. An agent closes loops.
Photo: Steve A Johnson, CC BY 2.0
The test: who does the follow-through?
Ask of any "AI PM feature": after it produces text, who acts on it?
- A summary of overdue tasks that a human must read, interpret and chase → chatbot.
- The same overdue list where each item gets a reminder sent, an escalation scheduled, and a note back to the client thread → agent.
The value was never the text. It's the follow-through — the six reliable workflows in 2026 are all follow-through: status rollups, tasks from notes, dependency checks, risk flags, report drafts, week summaries.
Checkpoints are the design decision
The mistake isn't giving agents too little autonomy — it's drawing the checkpoint line in the wrong place. Routine, reversible, provable-from-the-board actions (answering "what's the status?") shouldn't wait for a human. Commitments — scope, dates, money — always should. We wrote about that split in human-in-the-loop client communication.
What this looks like in agency work
The agent sits in the client's chat and on the board. It answers what it can prove, checks new asks against the scope of work, chases deadlines, and flags to a person anything that smells like commitment or conflict. The PM's job doesn't disappear — it compresses into the checkpoints, which is the part that was worth paying a PM for anyway.
That's the working definition behind Gavril: agentic where the work is mechanical, human where it's judgment.
← All posts