BlogsSep 12, 2026
When AI agents become abundant, new ways of working become possible
The economics of agentic work formations
By Walid Negm
AI agents are becoming capable of reasoning, planning, using tools, coordinating and acting with increasing autonomy. Their reliability remains uneven, particularly across long and messy work, but the economic shift is already visible.
The obvious opportunity is to use agents to perform more of today’s work. The larger opportunity is to reconsider how the work itself comes together.
Every way of working has always had to answer two questions:
How does the work progress? What makes the result official?
And underneath the first, where the structure does not already answer it: What gives participants the go-ahead to act? On a set path, finishing one step starts the next and there is nothing more to say. On a shared board, different participants act under different conditions, and that is most of what there is to know.
Take a simple refund. The process might carry the coordination: verify the purchase, check the policy, calculate the refund. A customer-service manager or a policy may give a specialist the go-ahead to approve an exception. And an authorized transaction in the payment system makes the result official.
Change those answers and you change how the work operates. An agent might coordinate an exception, a specialist might be granted temporary authority, or several independent judgments might be required before settlement.
Combine those answers with the people, agents and systems doing the work, and you have a work Formation.
AI did not invent Formation. It changes the economics of what we can form.
The economics of how work is organized
A refund does not only cost the work of issuing it. It costs getting the work to happen, holding the case together, checking what must be checked before the money moves, and bearing the consequence if the result is wrong.
A useful way to express that is:
doing it + holding it together + checking it + chance of being wrong × what wrong costs
C = E + Ce + R + pL
| Term | What it includes | Refund example | |
|---|---|---|---|
| doing it | E — Execution | Labor, agent and implementation cost, per role and per task. | Look up the order, apply the policy, write the credit. |
| holding it together | Cₑ — Coordination | Context assembly, model calls, duplicated reasoning, latency, reconciliation, and the cost of keeping participants aligned. | The agent pulling policy, order and payment state into one context; the extra calls if a second agent “helps”; the wait if they disagree. |
| checking it | R — Review | Review and reconciliation effort — a direct function of what settles, and of how many participants must agree before it does. | A lead reviewing an exception; two reviewers if fraud is suspected; reconciliation after the credit posts. Review is not “being careful” — it is a function of what must be true before the payment is allowed to settle. |
| being wrong | p×L — Failure exposure | Probability times consequence. The term that decides whether a faster arrangement is actually a better one. | Paying a refund that should have been denied — cash, chargeback, a fraud ring — or denying one that should have been paid — the customer, the chargeback fight, the regulator. |
A group of agents can make doing itlook cheap while making the overall work more expensive. More participants can mean more context, calls, waiting, reconciliation and review — and sometimes a more expensive mistake.
If you only count execution, you are not comparing arrangements. You are comparing one part of their bill. An arrangement-changing proposal should not be declared economically superior from execution savings alone; where it creates coordination, review or failure exposure the model cannot yet measure, the answer is not zero. It is that the economic comparison is incomplete.
The coordination cost function is moving
Organizations have historically designed work around two scarce resources: human attention and human judgment. That naturally favors defined processes and management. If the next action can be encoded in a process, nobody has to continuously watch and decide what should happen next. Where it cannot, a manager can make the call.
Consider a shared operational board. A shipment becomes late, inventory falls below plan and a customer commitment becomes exposed. Logistics, procurement, production and customer service can see the same changing state and respond when something relevant to their role occurs. With people, that can require several people to remain attentive, interpret what changed and decide whether to act.
Or consider a peer network. One participant raises an issue, another contributes evidence, another challenges the conclusion and another revises it until an explicit convergence rule is satisfied. Doing that with people for every case is expensive.
Organizations have therefore often simplified richer forms of coordination into alerts, queues, assignments, meetings and handoffs.
Reasoning AI changes those constraints. AI agents can remain available, observe changing conditions, interpret them, reason within stated boundaries and coordinate with others at a very different cost.
But coordination does not become free. The constraints shift toward context, computation, latency, verification, reconciliation, control and trust. Ten agents can still generate more work than they absorb. The scarcity moves.
A shared board, market, ledger or peer network can therefore become a viable design choice where previously it may have been uneconomic. Changing who performs the work and changing how the work is organized are economically different decisions.
The economics follow the work
The economic terms connect directly to how the work operates.
| Work-system question | What it can change economically |
|---|---|
| Who does the work? | Execution cost, capacity and throughput |
| How does the work progress? | Execution and coordination cost |
| What gives participants the go-ahead to act? | Monitoring, waiting, handoffs, contention and unnecessary activation |
| What makes the result official? | Review, control and commitment cost |
| What happens when it is wrong? | Failure exposure |
Review is not simply “being careful.” It follows from what must be true before the result can become official, including required evidence, agreement and reversibility.
Changing an arrangement can move several terms at once. A five-agent team may reduce human execution cost while increasing coordination, latency and review. One capable agent may cost less overall. Conventional software may cost less still. A deterministic route may win when the work is predictable or the consequences of failure are high.
We can now design work systems that operate differently
Work can progress through three broad coordination carriers: a known route, a directing participant, or an environment and rules through which participants coordinate without someone directing every move.
But the carrier alone does not define the arrangement. A handoff, market, committee, shared board and peer network can all use an environment and rules while differing in what gives participants the go-ahead to act and what makes the result official.
| Arrangement | How does the work progress? | What gives participants the go-ahead? | What makes the result official? | Example |
|---|---|---|---|---|
| Workflowsequence · decision · DAG | Route | the authored route — completing a step activates the next | last step completes | Claims processing |
| Manager → specialistsplanner · supervisor · lead | Directing participant | the lead, which assigns and redirects | the lead accepts or synthesizes the combined result | Insurance claims review |
| Handoffrelay · ownership transfer | Environment + rules | current ownership, under the transfer rules | valid transfer plus established assumption of ownership | Clinical handoff |
| Task marketcontract net · bidding | Environment + rules | the allocation rule, meeting bidder choices | the award is committed | Task market |
| Committeeindependent judges · panel · ensemble | Environment + rules | participation and eligibility rules — who may submit a judgment | aggregation rule produces a decision | Investment committee |
| Shared boardblackboard · shared workspace | Environment + rules | shared state plus activation rules; no director | the resulting state transition is validly committed | Inventory exception management |
| Governed meshpeer network · group chat · swarm | Environment + rules | the peer protocol's interaction rights | protocol determines convergence | Fraud network |
These arrangements are not new. Some are decades old. What changes is the feasibility of putting increasingly capable software participants inside them, combining them at finer granularity, and using them where continuous human attention and judgment once made them uneconomic.
A swarm, for example, is not itself an arrangement — it names the participants rather than the coordination. It tells us that several autonomous agents take part, but not how they coordinate, when they may act or what makes their collective result authoritative. Specify those rules and the swarm becomes something concrete: perhaps a governed mesh, shared board or handoff.
Nor is environment-and-rule coordination an AI invention. Kanban is a familiar example: shared state and pull rules let participants coordinate without a manager assigning every next action.
AI and agentic operating models open a larger design space for work.
But possible is not the same as operable.
A different arrangement can change authority, policy, risk, review, accountability and failure modes. If work moves from a fixed route to an agent-directed team or shared environment, the organization has to know who may act, what authority they hold, what requires review, what happens when something fails, and what makes the result official.
The same applies if work changes arrangement while it is running. That is not a routing decision — state, responsibility, authority and settlement cannot simply disappear at the boundary, and two formations do not unknowingly act on the same state. The transition itself is governed work.
What a governed transition has to state
- Why change?
- What condition requires different coordination?
- Who authorizes it?
- Who may change how the work operates?
- What carries forward?
- State · evidence · history · obligations
- What happens to work already in flight?
- Complete · cancel · transfer · reconcile
- What authority changes?
- Grant · retain · revoke · expire
- What must settle?
- What must be true before the old Formation ends?
Unstated, the transition still happens — it just happens as the residue of the first implementation, and unpicking that later costs a project.
And capability does not settle those questions.
“An agent could do this” and “an agent may do this” are different questions.
A recommendation to pay a claim is not the payment. A diagnosis is not a discharge. A model’s confidence does not confer organizational authority — probability is not authority, and confidence is a property of the model rather than a grant from the organization.
A newly possible arrangement therefore still has to clear a harder test:
Can it operate reliably? Can we understand its economic and operational consequences? Does the value justify the cost, risk, governance and organizational change?
Sometimes a deterministic route will still win. Sometimes an agent-directed team, shared board or peer network will make sense where it did not before. The conceivable Formation space is larger than the production-ready Formation space.
That is the opportunity created by abundant AI agents.
The objective is not to make work as agentic as possible. It is to make how work is organized a deliberate choice again.