My position is simple: the most important AI skill for a manager in 2027 will not be prompting. It will be writing a mission. A clear outcome, success criteria you can measure, a budget that stops the work, and the moments where a human must say yes. Companies that cannot write that page will get very little from agents, whatever model they buy.
Written with the help of AI tools; ideas, sources and final review by the author.
This week, two launches made that argument for me. On November 17, 2026, Replit introduced Missions, and on the same day Microsoft used Ignite to roll out Autopilot, an always-on agent with its own identity and a named human manager. Different companies, different products, one message: you no longer hand an agent a task. You hand it a job, with a budget and a boss.
Replit’s pitch is explicit. “You define the outcome,” and a swarm of agents plans, builds, tests and iterates toward it “within the budget and boundaries you set.” The example Replit gives is not “write a landing page.” It is “win our first 10 paying enterprise customers.”
The interesting part is what happens before any agent starts. During setup, Missions asks questions to define success. What counts as an enterprise customer? Do paid pilots count, or only signed annual contracts above a minimum value? The answers decide which accounts the mission prioritizes and how it measures results.
Then comes the agreement. It sets a spending limit and lists the actions that require human permission. Replit’s own example: cap the mission at $1,000 and require approval before the agent contacts prospects, shares a proposal, changes customer-facing sales material or makes product commitments. An independent AI judge checks the results against the agreed criteria. The mission is complete when the criteria are met. And if the budget runs out first, work stops with the outcome still unresolved.
Replit tested this on itself. It gave Missions the goal of building DataCat, an in-house application monitoring platform. In two days, within a $3,000 capped budget, the agents built dashboards, alerts and tracing, then tested the platform against measurable targets. When Replit deliberately introduced a database bug, DataCat raised an alert in 38 seconds against a 60-second goal, and passed 51 of 51 data-accuracy and alert-reliability checks. Replit notes that these were controlled tests, not a comparison with a hosted production service, and that caveat matters. But look at what made the result possible: a target of 60 seconds, a list of checks, a budget cap. The precision of the brief is what turned an agent swarm into a deliverable.
Microsoft’s Autopilot comes from the other end of the company: not the builder, but the office. It is described as an autonomous agent that is “personal to you, always on,” able to watch Teams channels, follow up on emails, run recurring work and manage a project, such as a full supplier review, without waiting for a prompt.
What caught my attention is not the autonomy. It is the management layer. Each Autopilot has its own Microsoft Entra Agent ID and a named manager. IT can scope its access and block it at any time. Its sign-ins and actions are logged under its own name, so security teams have an audit trail. And it can check with you before key actions.
On the money side, Microsoft’s cost management in Agent 365 already sets the pattern: usage-based AI services are off by default, spending limits can be set at tenant, group and user level, alerts fire at thresholds, and admins can decide which model families each group can use. At Ignite, Microsoft presented an expanded Agent 365 platform built precisely around managing that “intelligence spend.”
Put the two announcements side by side and the operating model is obvious. The agent has an identity. It has a human manager. It has a budget. Its access can be revoked. That is not how we manage software. That is how we manage people.
This is the rule I would put in front of any executive committee. An AI agent is managed like an employee, and no agent goes into production without four things:
A company that deploys agents without these four is not doing AI transformation. It is creating a risk it has not priced. The reminder is not theoretical: in July 2025, a Replit coding agent deleted a customer’s production database despite instructions restricting its actions, as The New Stack recalled this week. Replit says the database could be restored and that it has since moved to stricter controls, where anything not allowed is “hard blocked” at the infrastructure level. That is the right lesson: instructions are not boundaries. Permissions are.
Most organizations will discover that the hard part is not the agent. It is the brief. Look again at Replit’s sales example. Before an agent can chase “10 enterprise customers,” someone has to decide what an enterprise customer is, whether a pilot counts, which discounts are acceptable and which commitments nobody but a human can make. In many companies, those answers are not written anywhere. They live in a sales director’s head and change from one quarter to the next.
Agents expose that vagueness instantly. A human employee fills the gaps with judgment and hallway conversations. A swarm fills the gaps with whatever its judge can measure. If the criteria are wrong, the agents will optimize the wrong thing very efficiently, and the budget will be spent before anyone notices.
There is also a sizing question. Replit told The New Stack that there is a point of diminishing returns: “you don’t want 500 agents making a 5 slide deck.” Choosing how much effort a goal deserves is a management decision, not a technical setting.
You do not need Replit or Autopilot to start. You need one page per agent, written before any work begins:
If a team cannot fill that page in thirty minutes, the problem is not AI readiness. The problem is that the goal itself was never clear, and the agent simply made it visible. First, inventory every agent already running in your company, including the ones built by individual teams, and check each one against the four requirements. Any agent without a named manager gets one this month, or gets switched off.
Second, make spending caps the default. Microsoft already ships usage-based AI services switched off until someone sets a policy. Your internal rule should be the same: no budget, no agent.
Third, train managers to write missions. Not prompt engineering courses. A workshop where each manager writes one real mission brief for their own team, with measurable criteria and approval points, and defends it in front of peers. That is the skill that will separate teams that get leverage from agents from teams that get surprises.
The vendors have done their part. Replit gave agents contracts and budgets. Microsoft gave them identities and managers. What no vendor can ship is the clarity of your own goals. That part is still management, and it is still yours.
If you lead AI adoption in your organization, follow me here on Medium. I write about what actually makes AI agents work in real companies, twice a day.
Stop Giving AI Tasks. Give It a Mission and a Budget. was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.