Train a worker

The Library grounds any worker on demand; the Academy is how you make a body of knowledge part of one specific worker for good.

What you’ll do

1. Open a worker’s Academy from its profile.
2. Train it on the material you want it to own.
3. See the training reflected as always-on knowledge and a certification.

1. Find the Academy

Open a worker’s profile and look at the Journey card. It links to that worker’s stops in 🎓 Academy (training and latest certification) and 🧪 Testing (the rubric bench). The Academy is the worker’s own school — distinct from this learning site.

2. Train it

Run the worker through training on the material it should master for its role. Unlike RAG — which fetches Library items only when a task happens to match — trained knowledge becomes always-on for that worker: it carries the material into every task without needing to retrieve it first.

Three ways knowledge sticks

• Library + RAG — fetched when relevant, available to any scoped worker.
• Academy training — always-on, baked into this worker.
• Skills & expertise — items in the worker’s profile codex, folded into its prompt.

Different tools for different jobs; the Academy is for the knowledge a role should never be without.

3. Certification

A completed Academy track leaves a certification — a versioned badge marking what the worker has been trained on, surfaced on its profile. You can pair it with 🧪 Testing to bench the worker against a rubric and confirm the training took.

Train the role, not the task

Reserve Academy training for durable, role-defining knowledge (“our claims-handling policy”), and leave one-off facts to the Library. That keeps each worker focused and its certification meaningful.

What’s worth training a worker on

Training costs effort, so spend it on knowledge a role can never afford to miss. Three good candidates:

Your regulated or high-consequence rules. If you work somewhere with rules that must apply to every answer — what you may claim about a product, what a quote must legally include, what you must never advise on — that belongs baked into the worker rather than retrieved on a hopeful match. The whole point is that it applies even when the task didn’t mention it.

The domain vocabulary your business runs on. Your product names, your internal shorthand, the difference between a “job” and a “project” in your world. A worker that has this trained doesn’t produce plausible-sounding text using the wrong words — which is one of the fastest ways for AI-written output to look obviously AI-written to a customer.

The judgement calls in one role. For a support-shaped worker: what earns a refund, what gets escalated, what’s a goodwill gesture. Pair the training with 🧪 Testing to bench it against a rubric — that’s how you find out whether the training took before it meets a customer, rather than after.

Train the durable, retrieve the current

A good split: anything that changes yearly (policy, doctrine, role standards) is worth training; anything that changes monthly (prices, stock, availability) belongs in the Library where you can correct it in one place. Training something volatile just means it’s confidently out of date.

Recap

The Academy turns a body of material into always-on knowledge for a single worker, certified and testable — going beyond on-demand RAG. Next: the knowledge a worker writes itself, by saving reusable procedures as it works.

📖 Reference in the docs: Workers

➡️ Next: Skills workers learn on the job