NGT Skills
The know-how agents apply instead of inventing it: what an NGT Skill is, how an agent uses one, and how skills sit beside instructions, tools and context.
An NGT Skill is a piece of operational know-how, written by NGT in the form an agent reads when its task calls for it: how to review the NAVs an administrator struck, how to issue a capital call, what a CSSF return needs, how to build a workflow in SPARK. SPARK ships a library of about 500. They are the third of the five layers every use case is built from: NGT Skills own the procedure, and NGT Tools do the work (see What is SPARK?).
Why skills
A model that knows finance in general still has to guess how an operation actually runs, and guessing is where models invent. A skill takes the guess away: the agent applies a procedure whose checks are already in order and whose tools are already named, instead of working one out.
Take Daily Share Class NAV Oversight Review. It tells an agent to:
- Agree the NAV date and the funds in scope.
- Check that the Share Class Evolution results for that date are there, and if they are not, find out why the workflow that computes them did not deliver.
- List the share classes outside their expected range.
- Explain each from the data: sister classes moving together point to the portfolio; one class alone points to a dividend, an FX hedge or a fee accrual.
- Record what it found.
It also says what not to do: never recompute a NAV, and never write an explanation the records do not support. The skill owns the procedure; the numbers come from NGT Tools.
Skills beside instructions, tools and context
An agent answers from four things, each with its own job:
| What it gives the agent | For example | |
|---|---|---|
| Instructions | Who this agent is, what it does and what it must not do. Written for one agent. | Review struck NAVs against tolerance; never strike one. |
| Skills | How to do a kind of task well. Shared by every agent that holds them, and read when the task calls for it. | Daily Share Class NAV Oversight Review |
| Tools | What the agent can do: read records, open a task, run a workflow, search the web. | Spark Create Task |
| Context | What this thread is about: the items it holds. | Today's NAV canvas and the breach task |
See Agents for the agent itself, and Thread context for the items a thread holds.
How an agent uses its skills
- It holds a set. An agent holds the skills attached to it in Agent Studio, and a few its tools bring along. An agent made from a template starts with the template's.
- It sees the list and reads what applies. On every turn the agent is given the name of each skill it holds and the line that says when to use it, not the whole skill. When one applies to the task, it reads it in full.
- It asks before going further. Skills point to each other. When an agent wants one it does not hold, the thread asks you first.
- You see what it read. Every skill the agent read shows in its answer's steps, and opens as the agent read it.
NGT Skills and your own
NGT Skills are the same in every workspace and read-only. NGT updates them with SPARK's releases, and each shows its Version. Your workspace adds its own custom skills beside them, written by people or by agents, in the same form.
To fit an NGT Skill to your operation, such as your tolerances or your escalation path, duplicate it and change the copy. The copy is yours, and it stays as you leave it when NGT updates the original.