Prompt engineering grew up and became infrastructure

Enterprise AI Management · August 2026 · PolyQ Team

Prompt engineering grew up and became infrastructure

The 2026 consensus is that writing a clever prompt is the least interesting part of the job. What matters is the system that decides what a model can see, proves a change is an improvement, and records who approved it.

“Context engineering” is the term that stuck, and the shift behind it is real. At any serious scale the question stops being how a prompt is worded and becomes which retrieved documents, tools, policies, and history are assembled into the window before the model ever runs — and who is allowed to change that.

  • The emerging enterprise pattern: Practitioner guidance converges on the same sequence: put every prompt in one registry, version it natively, apply role-based access control, gate every change behind automated evals, then lift hardcoded business logic out of application code into a governed context layer that any agent can read at runtime.
  • Fetch at runtime, not at deploy: Agents retrieving the approved version of a prompt through an API means a wording fix ships without a release. That is convenient right up until it is terrifying, which is exactly why the approval workflow and the audit trail have to exist first.
  • Evals are the gate, not the report: The meaningful change is evals moving from a thing you run afterwards to the thing that blocks a merge. Faithfulness, hallucination rate, task success, and trajectory quality are the metrics doing that work, and keeping versioned prompts alongside evaluator configs and test reports is what makes an audit survivable.
  • Context as infrastructure, not a prompt file: The guidance we agree with most is to treat context assembly as a pipeline with data curation, privacy controls, and a log of exactly which tokens went into each answer. Without that last part, no one can explain an output after the fact.
  • Where we sit: This is the thesis Enterprise AI Management was built on, and it is why the same discipline runs underneath every venture we operate rather than living in one team’s repository.

The industry spent two years treating prompts as prose and is now treating them as configuration under change control. That is the right direction, and it arrived a good deal later than it should have.

If your prompts live in three repositories and a shared document, the first useful step is inventory, not tooling. We are happy to walk through how we did it.

Sources

More from Thinking