Worked use cases¶
These are end-to-end exercises, not a list of things an agent might someday do. Each gives prerequisites, sample input, a checkable result, and failure cases. All sample people, policies, and records are fictional.
Start with one read-only example. Add scheduling, memory, or write access only after its manual version behaves correctly.
| Use case | Requires | Writes outside Soulacy? | Walkthrough |
|---|---|---|---|
| Notes → action plan | One working model | No | Beginner exercise |
| Handbook questions with sources | Embedding provider and knowledge base | No | Grounded answers |
| Briefing on an iPhone | Paired phone and mobile destination | Delivers a message | Generation → delivery → push |
| Remember a report format | Learning-enabled agent and reviewer | Saves an approved lesson locally | Teach, approve, test, disable |
| Reassign an account, then undo | Configured conditional-write endpoint | Yes, only reviewed fields | Account handoff |
| Promote an agent revision | Autopilot and explicit checks | Changes future routing | Verified release |
Choose the right mechanism¶
- “Use the facts in these documents.” Use a knowledge base, not a permanent personal preference containing the whole handbook.
- “Use this format in future reports.” Use an approved Notebook preference.
- “Run this every morning.” Use a schedule after a successful manual run.
- “Show me what will change before writing.” Use Safe Undo only if your integration meets its conditional-write requirements.
- “Show me evidence before promoting an agent.” Use mission checks and candidate runs. A formatting check does not verify real-world truth.
Common acceptance checklist¶
Before using real data, try normal, missing, and contradictory input, a denied permission, and an unavailable dependency. Confirm the result and the run record. Never retry a possibly completed external write blindly. The first-checks guide explains where to look.