Snapshot Only
This docs page is a non-interactive snapshot of the demo UI system.

Inputs and buttons are intentionally disabled here. Run the standalone demo locally from docs/demo to create or attach to real research runs against a live API instance.

Read the API reference · cd docs/demo && npm install && npm run dev

Standalone demo preview

See the research-run cockpit before you run it locally.

This page mirrors the standalone demo layout: run creation, current status, live workflow timeline, approvals, graph summaries, logs, and export preview.

Snapshot target http://localhost:8080/v1

The real standalone demo persists URL state and connects directly to your running API server.

Create run

Research run request

POST /v1/research-runs
or
Current run

Live status summary

SSE connected
Run ID 071521de
awaiting_approval
Phase awaiting_approval
Questions 14
Claims 8
Sources 11
Title

ADK tool governance deep research demo

Primary question

How should a Google ADK-based research assistant govern tool usage safely in an enterprise environment?

Report preview

Preliminary findings indicate that tool governance should combine identity propagation, bounded tool access, approval gates for external writes, and durable audit trails.

Polling every 3s

Run paused for approval. This snapshot shows the operator review state.

Timeline

Live workflow events

4 events
approval.requested Jun 22, 10:11:42

Scope proposal paused for gate A approval

Payload 
{
  "gate": "A",
  "required": true,
  "risk_level": "high",
  "message": "Scope proposal requires human approval before continuing."
}
question.created Jun 22, 10:11:08

Question graph version 2 created for governance and risk framing

Payload 
{
  "questions_version": 2,
  "count": 14,
  "perspective": "governance"
}
source.accepted Jun 22, 10:10:41

Accepted primary source for policy and audit requirements

Payload 
{
  "source_id": "src-013",
  "title": "Enterprise agent governance controls",
  "source_type": "primary"
}
node.completed Jun 22, 10:10:17

Completed perspective generation

Payload 
{
  "node_name": "perspective_generate",
  "duration_ms": 891
}
Approvals

Gate decisions

Gate A

Scope review

pending

Awaiting human confirmation that the risk framing and exclusions are appropriate.

Gate B

Research plan

not_required

This gate has not been reached yet in the current snapshot.

Research graph

Claims and sources

6 records
Claims 3
Sources 3
Recent claims
  • High-risk tool categories need explicit human approval gates before external side effects.
  • Identity propagation should be attached to every tool call for policy and audit enforcement.
  • Append-only audit logs materially improve post-run reviewability for enterprise adoption.
Recent sources
  • Enterprise agent governance controls
  • Internal platform guardrail policy v2
  • Model tool access review checklist
Concept map

Topic projection

1 version
Nodes 4
Edges 6
Governance · perspectiveApproval Gate Design · conceptAudit and Traceability · conceptPolicy Enforcement · concept
Run logs

File-backed records

2 rows
INFO Jun 22, 10:11:42

approval requested for gate A

Record 
{
  "level": "INFO",
  "logger": "deep_research.workflow.graph",
  "message": "approval requested for gate A",
  "run_id": "071521de",
  "phase": "awaiting_approval"
}
INFO Jun 22, 10:11:08

question graph version 2 committed

Record 
{
  "level": "INFO",
  "logger": "deep_research.workflow.graph",
  "message": "question graph version 2 committed",
  "run_id": "071521de",
  "question_count": 14
}
Export

Report export preview

Format markdown
Content length 548

Snapshot data only. The live standalone demo returns export content from your running API.

# ADK tool governance deep research demo

## Executive summary

The current snapshot suggests a production research runtime should separate read-only retrieval from write-capable tools, require human approval for high-risk external actions, and carry principal identity through every workflow node.

## Working recommendations

1. Preserve identity context on every tool invocation.
2. Enforce explicit approval gates for external side effects.
3. Keep evidence, claims, and audit events independently inspectable.
4. Prefer durable checkpoints so interrupted runs can resume safely.