You can believe the ferry is coming… but keep the bridge until you can see it.
That is the lesson I keep taking from Reuters’ reporting on Meta’s Project OT—not that AI agents failed, and not that AI will never change the shape of an organization.
The mistake was redesigning the organization around a capability that had not yet proved it could carry the load.
Reuters reported in August that Project OT explored making Meta more “AI native” by using agents to perform work now done by employees. Some scenarios would have reduced certain teams by as much as 60 percent, collapsed roles into more generalized builder positions, and left smaller groups of people supervising virtual workers.
That is not an absurd future and Meta is still moving toward it. In August, Mark Zuckerberg described a future in which people use personal agents to build businesses and small groups can operate at significant scale. On September 8, Meta announced Muse, an agent that can browse, fill out forms, coordinate longer-running work, and ask for approval before sensitive actions.
So the technology did not stop, but direction is not arrival.
Reuters also reported that the organizational plan ran ahead of the evidence. Employees could produce substantially more code, but downstream impact did not increase at the same rate. Reliability and coordination problems grew. A larger second wave of planned cuts was called off, and Zuckerberg later acknowledged that agent development had not accelerated as quickly as expected.
That sequence matters.
Capability Should Precede Dependency
A responsible transition normally moves through stages:
- A capability appears.
- Teams experiment with it.
- They learn where it works and where it fails.
- Controls and recovery paths develop.
- Operating practices mature.
- The organization increases its dependence on the capability.
AI enthusiasm has a habit of compressing those stages into one executive sentence: We are becoming AI native. An aspiration is not an operating model.
If the technology is still experimental, the organization around it should preserve room for the experiment to be wrong. You do not remove the bridge because engineers expect the ferry to become reliable next quarter.
This is especially important because the two systems move at different speeds—the model can be updated next Tuesday, but the organization cannot.
Lay off employees and institutional knowledge leaves with them. Eliminate specialist roles and expertise disperses. Flatten management and coordination paths change. Move people from doing the work to supervising systems that are not yet reliable, and the old capacity can decay before the replacement capacity exists.
Technology changes quickly. Human systems change slowly and carry history with them.
If a company redesigns the slow-moving system around optimistic assumptions about the fast-moving one, it creates fragility that can survive the technical experiment that caused it.
The Org Chart Is Not the Dependency Graph
There is a second mistake hiding inside the phrase “AI native”: treating it as a synonym for startup-shaped.
Tiny teams, flat structures, fluid roles, and rapid iteration can work in startups because the enabling conditions are different. They may have narrow scope, high shared context, concentrated authority, younger infrastructure, fewer legacy obligations, and a tolerance for failure that does not map cleanly onto a mature enterprise.
A large organization can copy the visible topology without copying those conditions.
A three-person pod may still depend on security, legal, privacy, platform, data, deployment, operations, documentation, and other teams.
The org chart may be flat. The dependency graph is not.
Reducing the visible team does not remove the system around it. Often it moves coordination and verification work somewhere less visible.
That is why output metrics alone are dangerous. More generated code is not the same as more shipped value. If review, debugging, incident response, integration, and recovery costs rise downstream, the organization may simply have moved the work rather than removed it.
Meta’s Newer Agent Makes the Point Clearer
Meta’s September launch of Muse is an important update to this story.
The product is more capable than the systems described in the earlier reporting. Meta says Muse can take action across services, continue longer-running tasks, and use connected accounts. Meta also describes a dedicated secure virtual machine, a separate Sentinel agent that reviews outbound actions, scoped access, an audit trail, and human approval for sensitive actions.
Those controls are not side notes. They are part of the product.
And they reinforce the organizational lesson: useful autonomy depends on containment, observability, scoped authority, and explicit approval boundaries. A capable agent is not just a model. It is a model inside a control system.
Shipping that kind of consumer agent also does not prove that agents are ready to replace large portions of an internal workforce. Browsing for a person and carrying a company’s institutional responsibilities are different operating problems. The second includes accumulated context, cross-functional coordination, accountability, exception handling, and consequences that do not fit neatly into a benchmark.
The right question is not whether the agent can do impressive work.
It is whether the surrounding system can safely depend on that work.
Use Staged Dependency
The safer approach is staged dependency:
- Use agents before reorganizing around them.
- Measure completed outcomes, not only generated output.
- Increase authority gradually.
- Identify failure modes and downstream work.
- Build monitoring, approval, and recovery mechanisms.
- Preserve the human functions that detect weak signals and carry institutional context.
- Change the organization only after the capability and its control system have proved reliable under real conditions.
This may feel slower than declaring the company AI native but it is also how complex systems survive transitions.
The Meta story should not be read as evidence that AI will never transform organizations. Meta’s own releases since the original experiment make that interpretation harder to defend. It should be read as a warning against confusing a plausible destination with present-tense capacity.
You can believe the ferry is coming… but keep the bridge until it is carrying real traffic.