The question every team asks first is how much the agent is allowed to do. It is the right question with the wrong shape. Autonomy is not a single setting you turn up as trust grows. It is a ladder, and different kinds of work sit on different rungs at the same time. The ICSE 2026 workshop paper on agentic project management makes the same point formally: it proposes four working modes with varying autonomy, chosen per task, precisely because one level for everything cannot answer the accountability question.
1. Level 1: Read Only
The agent reads the workspace and writes nothing to it. It answers questions, summarises threads, drafts the standup and the Friday report for you to send. This is where every team should start and it already removes most status meetings. In Kavanah a read-only employee is one whose scopes contain no write tools at all; it cannot create a task even if asked nicely.
2. Level 2: Safe Mode
The agent can create and update work, plan a sprint, assign tasks and comment, but every external send, payment and delete is stripped from the run before it starts. Not refused at the moment of use, removed from the toolkit. Most teams keep triage and planning here permanently, because the cost of a wrong task assignment is a thirty-second fix and the cost of a wrong client email is a phone call. This is the default for a new AI employee.
3. Level 3: Approve
Whole categories of action park for a human before they run. External email, anything touching money, deletions, campaign sends. The agent prepares the action, the draft and the reasoning, and a person approves or edits it from a queue. The important property is that an approval is a gate, not a suggestion: the agent cannot route around it by choosing a different tool, because the policy is on the category. This is the right rung for client communication and invoicing.
4. Level 4: Full, Within Policy
The agent acts on its own inside the scopes, budgets and guardrails you set, and every action is written to a ledger with an undo where one is possible. This is for internal coordination at volume: standups, nudges, capacity rebalancing, the revenue forecast, scheduled reports. Nobody should approve a Tuesday standup summary by hand. The guardrails still hold: a per-tool block means the COO agent can move tasks all day and never archive a deal.
5. Putting Work on the Right Rung
Walk through the recurring work and ask two questions of each piece: how expensive is a wrong action, and how reversible is it? Cheap and reversible goes to level 4. Expensive but reversible goes to level 2. Anything that leaves the building goes to level 3. Anything you would not let a new hire do in week one stays at level 1. In Kavanah these are the autonomy level, the action policies and the guardrails on each AI employee, so a workspace typically runs a level-4 standup agent, a level-2 planner and a level-3 client desk at the same time.
The ladder is why the human project manager's job gets denser rather than smaller. Deciding which rung each kind of work belongs on, and reading the ledger to check, is the judgment part of the role. It is also the part that was always the real job.
