Platform engineering for agents
If you want AI to deliver at scale, the question stops being which assistant to buy and becomes what platform agents run on. Here is what that platform needs.
A single engineer using an AI assistant needs almost no platform: an editor, a key, a repository. The moment you want several agents doing real work across a real codebase, that changes. Now you have a platform problem, and it looks a lot like the platform problems you already know.
The good news is that the discipline is familiar. The teams who run agents well are doing platform engineering, just with a non-human as the primary user.
What agents need that humans take for granted
A human engineer arrives with context, judgement, and a laptop already set up. An agent arrives with none of that, every time. The platform has to supply what the human brought for free.
Somewhere safe to run. An agent should be able to build, run, and test a change in an environment that is isolated from production and from other agents. Reproducible, disposable environments are the foundation. Without them, every agent is one bad command away from a mess, and you will not let it move fast because you cannot afford to.
Context to work from. An agent that does not know your conventions will invent its own. A maintained knowledge base, your standards, your decisions, the shape of your domain, is what lets every agent start from the team’s accumulated judgement instead of from zero. This is the highest-leverage and most-skipped piece.
Review it can be held to. Volume goes up; trust has to keep pace. That means review that is structured, consistent, and recorded: changes checked against your standards, an audit trail of who changed what and why, and a clear gate before anything reaches production.
It is the same platform, with one new user
Notice that none of this is exotic. Isolated environments, good documentation, strong review, audit trails: this is platform engineering as you already practise it. What is new is treating an agent as a first-class user of that platform, with the same needs as a human and a few sharper edges.
The sharper edges matter. Agents are tireless, fast, and entirely literal. They will follow a broken process precisely and at volume. A platform that was merely tolerable for humans, because humans route around its rough parts, becomes a real liability when an agent follows it exactly a hundred times a day.
Where to start
You do not need to build all of this before you begin. Start with isolated environments, because nothing else is safe without them. Then invest in the knowledge base, because it compounds: every improvement to it improves every future agent’s work. Review and orchestration come next, as volume grows enough to need them.
This is roughly the order we follow when we help teams build it, and it is the thinking behind Smithy, the open-source orchestrator we maintain. The tools matter less than the shape. Build the platform, and the agents become useful. Skip it, and you will spend your time cleaning up after them instead.