Agent Studio

Document ID: DASH-003
File Path: docs/10-dashboard/agent-studio.md
Version: 1.0.0
Status: Draft
Owner: AI Platform Team
Last Updated: 2026-06-27

1. Purpose

This document specifies Agent Studio — the workspace for designing, testing, observing, and publishing agents. It is the visual front end over the Agent Definition model and the Agents API.

2. Agent Designer

A form-driven editor for the agent definition:

Section	Configures
Identity	Name, description
Model	Model selector (capability/class/strategy) or pinned model
Instructions	System prompt / behavior
Tools	Attach tools from the Tools API (only those the project enables)
Memory	Memory scopes + toggles
Policies	Attach policies (e.g. PII guard)
Budget	Default cost/token ceilings

Tool and policy pickers show only resources the user is authorized to use.

3. Test Console

An interactive console to run the agent before publishing:

Input ─► :run (draft version) ─► live stream
                                   ├─ planner steps
                                   ├─ tool calls (inputs/outputs)
                                   ├─ model deltas
                                   └─ memory reads/writes

The console streams the run via Agents API §6, exposing each step so authors can see why the agent did what it did.

4. Trace & Step Inspector

For any run, the inspector shows the full execution trace:

Planner reasoning and chosen plan
Each tool invocation with arguments, result, duration, and cost
Each model call with the routing decision, tokens, and cost
Memory retrievals with score breakdowns

This makes agent behavior debuggable rather than opaque.

5. Sessions & Multi-Turn

Authors can test multi-turn behavior using sessions: the studio preserves conversation context and (optionally) sticky model routing across turns.

6. Evaluation (Planned Integration)

Run an agent against a set of test cases / golden outputs.
Compare versions side by side (quality, cost, latency).
Track regressions before publishing.

This ties into the planned Testing section and AI evaluation service.

7. Versioning & Publish

Edits create a draft; publish produces an immutable agent_version.
A version diff highlights changes to instructions, tools, model, and policies.
Running agents and sessions continue on their start version.

8. Cost & Usage Preview

Before publishing, the studio estimates per-run cost from the configured model and typical token usage (using LLM Gateway pricing), and shows actuals from test runs.

9. Templates

Start from agent templates (support agent, RAG assistant, code agent — see planned Examples) and customize.

10. Governance

Attaching tools/policies/memory respects the user’s scopes and project enablement.
Publishing requires agents:write; running tests requires agents:run.
All actions are audited via the API.

11. Dependencies

13. Revision History

Version	Date	Description
1.0.0	2026-06-27	Initial Agent Studio specification