Hire contractors, not employees: why A2A and ephemeral workers are the future of AI agents
An opinion. The winning agent architecture will not be one all-knowing assistant with standing access. It will be short-lived, single-purpose workers that talk to each other through open protocols.
Most companies picture an AI agent as a digital employee: one assistant that knows everything, has access to everything and runs all the time. I think that picture is wrong, and that it will prove expensive and unsafe. The future looks less like an employee and more like a bench of contractors. Each one is brought in for a single task with exactly the access it needs, and is gone when the job is done.
Two ideas make that possible: agent-to-agent protocols and ephemeral execution workers. I run my own agent systems this way, and I would not go back.
Ephemeral workers: born for a task, gone when it is done
An ephemeral worker is an agent that exists only for one job. It starts in a clean, isolated environment, receives a short-lived identity and the permissions that job requires, does the work, hands back the result and is destroyed. Nothing lingers.
That one design choice solves four problems at once.
- Security. A long-running agent with standing access to your email, files and systems is a standing target. If it is compromised, or simply confused, the damage is limited only by what it can reach. A worker that lives for ten minutes with access to one folder can do ten minutes of damage to one folder.
- Freshness. As I argued in an earlier piece, the model underneath your agent is out of date within months. A long-lived agent has to be migrated. An ephemeral worker is just started on the newer model the next time it is needed.
- Cost and scale. You pay for work, not for idle time. When a data room lands, you can start fifty workers in parallel, one per contract, and have none running an hour later.
- Audit. Every worker has a clear beginning, a mandate and an end. That is the record a regulator, an auditor or an investment committee will want to see.
What persists is the memory, not the worker. The knowledge sits in a store the company owns and is handed to each worker as the context its task needs. Workers come and go. The memory compounds.
A2A: how the contractors talk to each other
Short-lived specialists are only useful if they can work together. That is what agent-to-agent protocols provide. A2A, introduced by Google in 2025 and now an open project under the Linux Foundation, gives agents a standard way to describe what they can do, take on a task, report progress and return results, whoever built them and whichever model they run on.
It complements the Model Context Protocol. MCP connects an agent to tools and data. A2A connects agents to each other. Together they let you assemble a team from parts: an orchestrator that plans the work, a contract reader, a financial modeller, a reviewer that checks the others, each one replaceable on its own.
Open protocols also matter commercially. A company that builds on them can swap one provider’s agent for another’s without rebuilding everything around it. That is the same portability argument I made about memory, applied to the workers themselves.
Why the all-in-one assistant loses
The single super-assistant is appealing because it is simple to picture. But it concentrates risk, ages quickly and locks you in. It is the monolith of the AI era, and software has spent twenty years learning to break monoliths into small services that can be replaced, scaled and secured one at a time. Agents will follow the same path, only faster.
A familiar model
There is nothing new about the idea itself. It is how good organisations already use specialists. An interim executive is brought in with a clear mandate and the authority to deliver it, does the job and leaves the organisation stronger. Nobody gives a contractor the keys to every building forever.
We should hold our AI agents to the same standard: a clear mandate, the minimum access, a defined end and a full record of what they did. Hire contractors, not employees.