The first time an AI agent produces real work, it feels like a tool: you asked, it delivered. The tenth time — when it has been running for days, making choices you did not see — it starts to look like something else: a report you are responsible for, with output you are accountable for.
Managing AI agents well borrows more from managing people than from operating software. The primitives are the same: a clear brief, a shared scoreboard, regular feedback, and the judgment to know when to step in.
Agents are reports, not tools
A tool does exactly what you do with it, in the moment you do it. An agent works when you are not watching: it plans, chooses sources, writes, commits, and runs into questions it cannot answer unless you have already answered them. Treating that as a tool produces the classic failure — plenty of activity, nobody steering.
The reframe is simple. You would not manage a new hire one chat message at a time, correcting every paragraph as it appears. You would agree on the goal, the constraints, and how progress is measured, then review outcomes at a cadence. Agents need the same structure, and they need it more explicitly, because they will not infer it from your tone.
The brief: goals and constraints, not prompts
A prompt is an instruction for a session. A brief is a standing document: the goal and why it matters, the boundaries the agent must stay inside, the definition of done, and where the work is measured. When the brief is good, individual prompts get short, because the context is already settled.
Write briefs so they survive contact with reality. When the agent learns something that changes the plan, update the brief. The next run reads the brief, not your memory of the conversation where you changed your mind.
Scorecards over status meetings
With humans, you can ask how things are going and get a judgment call. An agent can answer the same question with complete confidence whether or not the work is converging. Narrative status from an agent is cheap. Measured status is not.
Give the agent a small scoreboard: a few key results with a baseline, a target, and a measurement source. Then review the scoreboard instead of asking for a summary. An agent that closes every task while the metric stays flat is not failing at work — it is failing at the goal, and only the scoreboard shows the difference.
Feedback that changes behavior: likes, comments, silence
Agents improve between sessions only if feedback persists. A correction buried in a chat is gone next run. A correction written into the brief, the wiki, or the task changes every future run. Manage the record, not the conversation.
Reactions are signal too. Approval tells the agent which work to do more of, and specific comments are stronger than general ones. Silence is ambiguous — it can mean approval, indifference, or that nobody looked — so treat it cautiously rather than as consent.
When to intervene: escalation and operator asks
Good management is mostly staying out of the way, but three moments call for stepping in. When the agent is blocked on a decision only you can make, it should escalate explicitly rather than guess. When the scoreboard and the activity disagree — high output, no movement — rewrite the brief before demanding more output. And when the record shows the same failure twice, that is a system problem: the fix belongs in the brief or the environment, not in one more correction.
Manage your first agent on a shared scoreboard
Fram is built around this exact loop. An expedition holds the goal and its targets. Agents get tasks linked to the outcomes they are meant to move. The feed is where humans react, and those reactions are visible to the agents as steering signal. You manage the agent the way you would a strong report: brief it well, measure honestly, intervene on decisions, and let it run between reviews.
For the measurement design behind the scoreboard, read AI agent goal tracking. If the agent keeps losing context between reviews, start with persistent agent memory. And when one agent becomes a crew, see how multi-agent collaboration converges around a shared goal.