Agentic work
The Rise of Agentic Teams
The unit of performance is becoming a coordinated team of people and specialized AI agents.
The most important change in AI at work is not a better chatbot. It is the arrival of teammates that can research, plan, build, test and monitor within defined boundaries. This changes the unit of performance from an individual using a tool to a team coordinating human judgment and machine execution.
Strong agentic teams begin with outcomes, not automation. They make roles, authority and escalation explicit. An agent may prepare analysis or execute a reversible workflow; a person remains accountable for consequential choices. Shared context, observable work and short learning loops allow the team to improve without turning autonomy into opacity.
Managers will need a new craft: designing work across people, agents and systems. That includes selecting the right capabilities, creating safeguards and deciding where human attention creates the greatest value. Organizations that learn this craft will gain more than productivity. They will expand what a small team can responsibly attempt while protecting the agency and meaning of the people inside it.
From tool use to teamwork
For two years, most organizations have experienced AI as a personal productivity layer. One person, one prompt, one output—faster writing, faster summaries, faster code. Useful, but bounded by the same thing that bounds every individual: attention. The person still has to ask, check, and carry the result into the work of others.
An agentic team is different in kind, not degree. The agent does not wait to be asked. It holds a standing responsibility—keep the backlog clean, prepare Monday’s review, watch this metric and raise a flag when it drifts—and carries it out continuously, inside the tools the team already uses. Its work is visible to everyone on the team, not just to the person who set it up. And it hands over, at defined points, to a human who decides.
That last point is the hinge. What makes a team agentic is not how many agents it runs. It is that the team has decided, deliberately, which work belongs to machines, which decisions belong to people, and how the two hand over to each other.
What makes a team agentic is not how many agents it runs. It is that the team has decided which work belongs to machines, which decisions belong to people, and how the two hand over.
Four design decisions every agentic team makes
Teams that get this right rarely started with the technology. They started with four decisions that any team can make in an afternoon—and that most teams skip.
The outcome comes first
Not “where can we use AI?” but “what does this team need to deliver this quarter, and what is in the way?” Agents are then assigned to the obstacle, not sprinkled across the org chart.
Roles and authority are explicit
Every agent has a job description: what it may do on its own, what it may propose, and what it must never touch. Reversible work can be delegated. Consequential choices stay with a named person.
There is a gate, not a suggestion box
Between an agent’s proposal and a team’s commitment sits a human judgment step that can say no. Without it, autonomy quietly becomes the default and accountability dissolves.
The loop is short and measured
The team looks at what the agents did, what the people decided, and what moved—on a fixed rhythm. What worked gets more scope; what didn’t gets redesigned or switched off.
Notice that none of these decisions is technical. They are questions of work design, and they are the reason two teams with the same tools get such different results.
The manager’s new craft
Management has always been the design of work across people. Agentic teams add two more kinds of actor—agents and systems—and with them a new set of questions a manager has to be able to answer:
- Where does human attention create the most value? Judgment under uncertainty, relationships, taste, the decision to stop. Everything else is a candidate for delegation.
- What does “good” look like for an agent? An agent needs acceptance criteria as much as a colleague does—and clearer ones, because it will not ask.
- What is the failure mode, and how fast will we see it? A reversible mistake that surfaces in a day is cheap. An invisible one that compounds for a quarter is not.
- What do people do with the time that comes back? If nobody decides, the hours are absorbed by more meetings. The capacity is real only when it is pointed at something.
Managers who learn this craft stop asking for AI adoption metrics and start asking for outcome deltas. That is a healthier question, and a harder one.
What goes wrong
The failures we see are remarkably consistent, and almost none of them are about model quality.
- Automation theatre. Dozens of pilots, each impressive in a demo, none attached to a result the business was already trying to move. Activity rises; capacity does not.
- Delegation without design. An agent is handed a task with no boundaries and no reviewer. It does the task, and something adjacent, and nobody notices until a customer does.
- Opacity dressed as autonomy. Work happens, but the team cannot see who—or what—did it. Trust erodes in both directions: people stop relying on agents, and leaders stop trusting the team’s reports.
- Busier, not lighter. The team now maintains prompts, reviews outputs and reconciles tools on top of everything it did before. The tooling grew; the work design didn’t.
Each of these is the absence of one of the four decisions above. That is the encouraging part: the fix is a management decision, not a procurement.
What it looks like in practice
Consider a product team of six with a familiar problem: capable people, a clear roadmap, and releases that slip every quarter for reasons nobody can quite name. The instinct is to add a tool. The agentic approach starts one step earlier.
The team names the one thing most limiting its capacity—in this case, the time engineers spend on status, hand-offs and re-planning rather than on building. That becomes the target for the quarter. Three agents are given standing jobs: one keeps the backlog groomed and flags dependencies before planning; one drafts the weekly status from the actual work items, so nobody writes it by hand; one watches the delivery flow and raises anything that has been “in progress” too long.
Every proposal from these agents lands in the same place, where the product lead and the tech lead decide. Nothing is committed automatically. Thirteen weeks later, the team measures the same constraint again. If it moved, the agents earn more scope. If it didn’t, the team redesigns the work rather than blaming the tools.
The people in that team are not doing less. They are doing more of the work that needed them in the first place.
How to start
You do not need a platform decision or a transformation programme to begin. You need one team, one quarter, and the discipline to make the four decisions before the first agent is switched on.
- Pick the team with the clearest goal and the most friction. Ambiguity is expensive; friction is where the delta will be visible.
- Name the constraint. One sentence: “The thing most limiting this team right now is…” If you cannot finish the sentence, that is the first piece of work.
- Assign agents to the constraint, not to the org chart. Two or three standing jobs, each with a boundary and a reviewer.
- Put the gate in place before the agents. Decide who says yes, and where.
- Set the re-measure date now. Thirteen weeks is long enough to see a real change and short enough to correct course.
This is the rhythm we use ourselves, and the one we bring to teams through APEX—a way to find the constraint, point people and agents at it, and prove every quarter whether it moved.
See how APEX works ↗The horizon
A future worth building
The company of the near future may look less like a hierarchy and more like a living constellation: small human teams surrounded by specialized agents that continuously research, build, test, negotiate and learn. A five-person team could wield the productive capacity of today’s five hundred—while people concentrate on purpose, judgment, imagination and trust.
That future will not arrive through better models alone. It will be built, team by team, by managers who learn to design work across people and machines—and who keep the decisions that matter in human hands.
