Files and knowledge
Files give an agent material it can read, create, and return to. Some files belong to the conversation happening now. Others become part of what the agent can use the next time you work together.
Session files support the work happening now
Suppose you upload this month's campaign spreadsheet and ask the marketing launch agent to explain why one audience performed better than another.
The spreadsheet belongs to that session. The agent can inspect it, create a chart, and revise the analysis with you. A new session does not start with that temporary working set.
Session files are useful for:
- Material you upload for one request.
- Drafts you are still shaping in the conversation.
- Downloads and intermediate results from a longer task.
- Finished files you want to take away, but do not need the agent to keep using.
Think of the session files as the desk for today's work.
Agent files carry useful context forward
Now suppose the launch agent writes an email that your team approves with only one small edit. You want the next launch to benefit from that example.
Save the email in the agent's files, alongside the brand voice guide and approved product notes. Those files belong to the agent and are available across its sessions.
Agent files are useful for:
- Brand, product, and policy reference material.
- Examples of work your team approved.
- Finished reports worth consulting later.
- Notes and indexes that help the agent find its way around the collection.
Think of the agent files as the shared shelf the agent returns to from one conversation to the next.
| File scope | Best for | Launch agent example |
|---|---|---|
| Session | The task in front of you | This month's campaign spreadsheet |
| Agent | Work and knowledge worth reusing | The brand guide and approved launch email |
Turn a useful result into reusable knowledge
An agent's knowledge does not need to begin as a large, carefully designed library. It can grow from the work you already review.
After a strong launch, keep the approved brief and email. After the agent uses an outdated claim, update the product notes. When the collection becomes harder to navigate, add an index that says what each folder contains.
Then give the agent a short direction in its instructions:
Before drafting launch copy, read the current product notes in
`agent-files/product/` and the approved examples in
`agent-files/examples/launches/`.
The instructions tell the agent where to look and when. The files hold the detail.
A folder can carry rules for the work inside it
Some guidance applies only to one area. Your paid-campaign folder, for example, may have a naming convention, a spending limit, and a reporting format that do not belong in the agent's main instructions.
You can place an AGENTS.md inside that folder. When the agent works there, the folder-level instructions give it the local rules alongside its main instructions.
This keeps broad working rules in one place and specialized rules close to the files they govern.
The agent reads files when the task needs them
Agent files are available to the agent, but their full contents do not need to sit in every conversation. The agent can open the relevant product note, example, or report when the task calls for it.
Clear names, a simple folder structure, and short pointers in the instructions make that choice easier. A file named enterprise-security-approved-claims.md is more useful than notes-final-2.md; an index is more useful than a folder the agent has to guess its way through.
Use a skill when the agent needs a repeatable process. Use files when it needs material to read or create. A launch-brief skill can tell the agent which product notes and examples to consult without copying all of that material into the skill itself.
Continue with the guide
Build a knowledge base for your agent shows how to save agent files, point to them from AGENTS.md, and organize a reusable collection.
Related concepts: