Agent Studio: build your own AI employee
Compose a custom agent with its own instructions, knowledge, and actions, then map it into Ask AI so your team can summon it by name.
- 1
Pick a job with clear edges
Good first agents do one describable job: a quoting assistant that prices from your rate card, an onboarding guide that walks new clients through setup, a content brief writer in your voice.
- 2
Write instructions like an SOP
In Agent Studio, the agent’s instructions are its employment contract: role, tone, what it always does, what it never does, and what it hands to a human. Specific beats clever.
- 3
Attach knowledge and actions
Upload the documents it reasons from (rate cards, playbooks, policies) and grant only the actions its job needs. Least privilege applies to AI employees too.
- 4
Test adversarially
Before rollout, try to break it: ambiguous requests, out-of-scope asks, missing information. Patch instructions where it flailed.
- 5
Map it into Ask AI
Connect the agent so it is callable from the Ask AI surface; the team summons it in-platform where the work already happens.
- 6
Version the instructions
Keep the instruction text in a doc with dated changes. When behavior drifts or improves, you want to know exactly which edit did it.
Do it with Claude
Claude drafts the agent instructions from your SOP for the role, generates the adversarial test scripts, and rewrites the instruction patch when a test fails.
Watch out for
- One agent asked to do five jobs, doing all five poorly.
- Granting broad actions to save setup time; scope creep in agents is a real risk surface.
