A pre-trained agent is an AI agent prepared in advance for a common job, such as reconciling bank statements or triaging complaints, with a starting brief, typical sources and behaviour already tested on that kind of work. It saves the build, not the proof. It still has to earn its authority on the customer's own cases before it acts.
What comes ready-made in a pre-trained agent?
The parts that are the same for most firms doing the job. A role statement covering what the job includes and excludes. A list of the sources the job normally needs. The connections it usually requires, such as read access to a ledger and write access to a case queue. And a body of test cases that show how the agent should handle routine work, edge cases and refusals.
What does not come ready-made is anything specific to you: your policies, your limits, your customers, your systems. Those are filled in when the agent is set up for a particular business.
Why does the distinction matter once agents act?
Vendors often present a pre-built agent as if it arrives approved. It cannot. Whether an agent handles your supplier disputes well depends on your suppliers, your contract terms and your thresholds. Testing on someone else's operation tells you the agent is a reasonable starting point. It does not tell you it is safe to act in yours.
So a pre-trained agent starts at read-only, like any new agent. It earns more depth once it has been replayed against your real cases and scored. Depth is earned, not configured.
How is a pre-trained agent set up in practice?
A team picks the agent closest to the job. They edit its agent brief to match their process, attach their own sources with the right provenance class, and connect it to their systems. Then it goes through certification on their cases. The agent earns the highest depth its cases prove; if it clears none, it stays at read-only until fixed.
Take an accounts payable agent at a regional distributor. The starting brief already covers matching invoices to purchase orders and flagging mismatches. The distributor adds its tolerance rules, its approved-supplier list and its own exception queue. Certification shows the agent matches reliably but mishandles credit notes, so matching runs at depth 3 while credit notes stay with a person. How the agent behaves when a step fails partway is tested too; see how AI agents recover from errors mid-task.
How is a pre-trained agent different from a pre-trained model?
A pre-trained model is a language model trained on a large general corpus before any task-specific use. "Pre-trained" there refers to model weights. A pre-trained agent is built on top of a model, and its preparation is in the brief, sources and test cases, not the weights. The same pre-trained agent can run on different models, and setting it up for a customer never changes a model's weights. We call this qualification, not training.
Where Pryme fits
Pryme's Pre-Trained Agents arrive with a brief, typical sources and test cases for common jobs. Each one is then qualified for the customer: the brief is edited, declared sources and others are attached with their class, and Attest replays the agent against the customer's real cases before the agent goes above read-only. See the governed agentic operating system, explained.
Questions people ask
Is a pre-trained agent faster to deploy than one built from scratch?
Usually, because the brief and test cases already exist. Certification on your own cases still has to happen, so the saving is in drafting and setup, not in proof.
Can a pre-trained agent's certificate carry over between customers?
No. A certificate covers one agent for one customer, proven on that customer's cases. A second customer runs its own certification.