Governed research runtime

Research that can be inspected, challenged, approved, and shipped.

Deep Research Assistant turns a vague research request into a bounded workflow with explicit scope, question graphs, evidence fragments, claim records, approval gates, and a final report that can be reviewed like engineering output instead of chat history.

Built on Google ADK 2.0 with deterministic workflow nodes and bounded Gemini-backed agents.
review_first run
Objective

Assess enterprise controls for an AI research runtime.

Mode

review_first

Evidence

Immutable excerpts with source and question linkage.

Claims

Atomic, qualified, and attached to evidence ids.

Approvals

Scope, plan, outline, and publication gates.

Output

Report draft plus progress, graph, events, and exports.

Why it exists

Most “research agents” produce prose. This one produces inspectable work products.

Core promise

Every material statement should trace back to evidence, not prompt theater.

Target use

Technical reviews, architecture analysis, governance-heavy research, and decision support.

Product features

Everything important in the workflow is first-class.

Evidence-first pipeline

Intent becomes scope, perspectives, questions, search plans, evidence fragments, claims, contradictions, drafts, verification findings, and report output. The workflow is explicit at every stage.

Bounded multi-agent design

Research Director, Evidence Curator, Claim Builder, Verifier, Moderator, and supporting agents each have narrow responsibilities instead of sharing one opaque prompt.

Approval-aware operation

Review gates are built into the graph, not bolted on later. Runs can pause at semantic boundaries before scope expansion, outline approval, or publication.

Routeable API surface

The service exposes create/get run endpoints plus graph, frontier, progress, events, concept map, approvals, interventions, and exports for integration into real systems.

Deterministic validation

Route contracts, workflow behavior, and quality gates run in normal CI. Live Gemini checks stay opt-in and bounded so cost and flake do not leak into every push.

Enterprise control plane

Identity propagation, policy checks, source filtering, budget accounting, telemetry, and append-only audit events are all part of the system design.

Workflow

From broad prompt to reviewable artifact.

01 Scope and objective proposal
02 Perspective and question graph generation
03 Search planning and source intake
04 Evidence extraction and claim construction
05 Contradiction checks and coverage review
06 Outline, drafting, verification, approval, export
For builders

Built to plug into product and ops surfaces, not just demos.

The project already documents and tests the API endpoints that matter for orchestration: run creation, run retrieval, graph inspection, frontier visibility, event replay, approval state, interventions, and export.

See the documented routes →
POST/v1/research-runs
GET/v1/research-runs/{run_id}
GET/graph /frontier /progress
GET/events /concept-map
POST/interventions /exports
GET/POST/approvals
Why this is different

It advertises rigor, not just autonomy.

Deep Research Assistant

  • Claims are explicit records
  • Evidence stays immutable
  • Workflow events are inspectable
  • Approvals are modeled in the graph
  • Validation is split between deterministic and live tiers

Typical agent demos

  • Prompt in, prose out
  • Weak traceability to sources
  • Little state visibility
  • Governance handled outside the workflow
  • Live dependencies mixed into normal CI
Next steps

Read the design, inspect the API, or validate the live path yourself.