Opinion
More AI agents can mean more work for humans
As companies add more autonomous tools, employees are increasingly becoming the ones who move context, coordinate handoffs and keep disconnected AI systems working together.
A strange thing is happening to some of the developers who are furthest ahead in adopting AI. They may have several coding agents running at once - one planning a task, another writing code, another reviewing it - yet a surprising amount of their own time is spent copying information between them. The agents are getting better, but the human is increasingly becoming the router: moving context, deciding what goes where and keeping separate sessions aligned.
The same coordination problem is beginning to appear beyond software development. Companies are adding autonomous AI tools across engineering, security and operations. Each agent can perform more work independently, yet these systems often operate as isolated participants rather than as a team. As their numbers grow, the productivity bottleneck shifts from doing the work to coordinating it. With one AI assistant, that overhead is barely noticeable. With five or ten agents, someone has to decide which system needs which piece of context, pass outputs from one to another, restart a process when a chat is compacted or the session restarts, and make sure the next participant understands what has already happened.
As agent output accelerates, human capacity becomes another bottleneck. An agent can generate, analyze or revise work far faster than a person can inspect every intermediate step; but the need for coordination and review becomes more important, not less. If every increase in machine output creates a larger queue of handoffs, checks and decisions for people, the organization has automated production while leaving the control flow largely manual.
The problem becomes even more obvious when agents start communicating directly. Connecting two intelligent systems is not the same as connecting two traditional software services. AI agents are probabilistic participants. Give two of them an open-ended conversation and they may continue responding after anything useful is being accomplished. Add several more and every message can trigger unnecessary responses, consume tokens or create new branches of activity. The challenge is therefore not simply to let agents talk, but to make sure the right participants receive the right context, at the right point in the task, and know when their task is finished.
Let’s say that we connected two coding agents and let them respond freely to one another. They can recognize that a conversation has become circular, but they can’t necessarily stop it. One may try to persuade the other to stop; the other responds, and the loop continues. The problem is not that either agent is incapable. It is that they are missing rules around routing, relevance, state and stopping conditions; and these do not emerge automatically from the intelligence of the AI agents themselves.
The natural response has been orchestration: define a workflow, decide which agent goes first and specify what happens next. That works well when the path is predictable. But many of the tasks we want agents to perform are valuable precisely because the path is not completely known in advance. A planning agent may discover that it needs help from a security agent. A coding agent may need to bring in a testing agent. One participant may fail halfway through a task and return later, while another may need context that did not exist when the workflow began. In those situations, the problem starts to resemble managing a team more than automating a checklist.
Humans have developed collaborative approaches to our work over time. We know how to behave when someone joins a conversation, how to provide them the context they need, who needs to be on which email threads, and when a discussion has gone far enough. We also understand roles and boundaries: the finance team does not need to be part of every engineering discussion, and a new employee should not immediately receive access to every system. AI agents do not have the same inherent knowledge that we do when it comes to teamwork. They need equivalents of our social mechanisms if organizations expect them to operate at scale. That means knowing which agents exist, who owns them, which systems they can access, who they are allowed to communicate with and what information they may exchange. It also means preserving context across handoffs and failures, keeping a record of what happened and giving humans the ability to inspect or stop actions when necessary.
None of this means removing people from the system. Human oversight becomes more valuable as agents gain the ability to act, especially around sensitive data, external systems or irreversible decisions. The distinction is between judgment and clerical coordination. A person should still be able to approve a high-risk action, resolve an ambiguous situation or intervene when the system behaves unexpectedly. They should not have to spend their day carrying context from one agent to another simply because those agents lack a shared way to coordinate. The goal is not autonomy without humans; it is to reserve the most valuable asset–us–for the decisions that actually require it.
These requirements become more important because enterprises are unlikely to live inside a single AI ecosystem. One team may use Claude, another Codex and another an internal agent, while business applications introduce agents of their own. As agent-based work expands across organizational boundaries, agents operated by partners and software providers may also need to participate in the same workflows. The result is not one centrally designed system but a distributed workforce of AI participants built by different vendors, running in different environments and operating under different owners. Direct collaboration therefore has to exist with identity, permissions and clear authority boundaries; otherwise, removing the human from the handoff simply removes the person who was also enforcing those boundaries.
If every interaction between those systems still requires a developer to copy, translate, approve and reconnect the pieces manually, the organization has not achieved much autonomy. It has created a new category of coordination work for humans. The first phase of the agent shift has been about making individual AI systems capable of doing more. The next phase is about making them capable of working together without turning the employee into permanent middleware. The real promise of autonomous agents will not be fulfilled when every employee has ten of them. It will be fulfilled when managing those ten agents no longer becomes a full-time job.
Arick Goomanovsky is Co-Founder and CEO of BAND.
