Deepin Harness

Run AI as an organization.

Deepin Harness is the runtime underneath every Deepin Investigation and Space. It gives agents the roles, capabilities, knowledge, policies and controls they need to work inside an enterprise.

Models reason. Capabilities act. Policies govern. Harness orchestrates.

Explore the architecture See it in action: deepin.risk →

Runs underneath deepin.risk and deepin.finance

Organization RoleAuthorityRoleAuthorityRoleAuthorityAgentAgentAgent Task Capability Tool · Service KnowledgePolicyMemoryEvaluation
  1. Organization
  2. Role · Authority
  3. Agent
  4. Task
  5. Capability
  6. Tool · Service
KnowledgePolicyMemoryEvaluation

The problem

Agents are easy. Organizations are hard.

A model with a few tools can already do useful work. The hard part starts when that agent has to work inside a company.

A common agent
ModelAgentTools

An agent can reason and call tools.

What an enterprise asks around it
  • Who is responsible?
  • What role does it have?
  • What is it allowed to do?
  • Which data can it access?
  • Who can delegate work to it?
  • Which capabilities can it use?
  • Which policies constrain it?
  • When is human approval required?
  • What happens when it fails?
  • How is its behavior evaluated?
  • Can every action be audited?

Enterprise AI is an organization problem.

Organization model

Give AI a place in the organization.

An enterprise agent should not be an anonymous model with tools. In Deepin Harness it works inside an explicit organizational context: what it is responsible for, what it may do, and under which rules.

OrganizationThe boundary in which roles, authority, knowledge, policies and capabilities exist.
Who, and under which rules
RoleWhat is it responsible for?
AuthorityWho may decide, approve, delegate or act?
Policy · ConstraintUnder what conditions may this happen?
Doing the work
AgentDoes the work, within its role.
TaskWhat needs to be done?
CapabilityWhat can it do?
ToolA technical way a capability gets done.API · MCP tool · query · model · function
ServiceInternal or external systems exposed to the organization.ERP · CRM · core banking · registry service · document system
What it knows
KnowledgeOrganizational information available to roles and agents.Documents · graphs · structured data · previous Investigations · organizational memory
A conceptual model, not a class diagram. Capability ≠ Tool: a capability is what can be done; a tool is one technical way to do it. Authority ≠ Policy: authority is who may decide, approve, delegate or act; policy is the rules under which it happens.

Organizational Reasoning

Reason within the organization.

Roles define responsibility. Authority defines delegation and decision rights. Capabilities define what can be done. Policies define the boundaries.

Organization

Structure and state: who exists, what they are responsible for, what they may do.

Organizational Reasoning

Behavior over time: how the organization works out who should do what when something happens.

Incoming
Organization boundary
Third-party riskagent
Registryagent
Dispute operationsagent
Investigation leadagent
Risk analystperson
Dispute analystperson
Compliance officerperson
Who is responsible?
Third-party risk role
What authority exists?
May investigate and propose. A person changes the risk level.
What can be done?
Registry lookup · relationship graph · change detection
What is allowed?
Supplier monitoring policy allows it
The organizationDelegates the registry check, then acts on the findings
New stateCase open · recommendation waiting for the risk analyst

Illustrative. Roles, authority and policies are examples; each organization defines its own.

Runtime

From model call to governed work.

Work enters the Harness, runs under organizational context and governance, and leaves an auditable trace.

Request or event
Deepin HarnessLeaves a trace
  1. 01Identity & rolewho
  2. 02Context & memorycontext
  3. 03Planningplan
  4. 04Capability selectioncapability
  5. 05Tool & service executioncall
  6. 06Policy enforcementpolicy
  7. 07Human approvalapproval
  8. 08State & workflowstate
  9. 09Evaluationscore
  10. 10Audit & observabilitytrace
Enterprise systems

Models reason. Capabilities act. Policies govern. Harness orchestrates.

A model call answers a prompt. Governed work knows who asked, in which role, with which capability, under which policy, and who approved it.

Governance

Autonomy needs boundaries.

Enterprise AI should be able to act. Every action still passes through role, authority, policy and constraints, and ends in one of three outcomes.

Pick a request
Action requestRead the latest registry filing
  1. Role
  2. Authority
  3. Policy
  4. Constraint
ExecuteHuman approvalDeny

Within the role, the authority and the monitoring policy. It runs, and the step is logged.

Controls
  • Role-based access
  • Authority boundaries
  • Policy enforcement
  • Human approval
  • Audit trail
  • Data lineage
  • On-premise
  • Air-gapped deployment

Controls are configured for each deployment, to the organization's own rules.

Knowledge & memory

An organization should remember.

Work should not start from zero every time. What the organization knows, and what it learned in earlier Investigations, is available to the roles doing the work.

  • Documents
  • Structured data
  • Knowledge graph
  • Previous Investigations
  • Policies
  • Decisions
  • Evaluation results
Organizational knowledgeShared context for the work
  • Roles
  • Agents
  • Investigations

Every Investigation can contribute reusable organizational context.

It stays yours.

This knowledge lives in your deployment and belongs to you. It is not shared between customers. What carries from one deployment to the next is how Investigations are solved, never your records.

Evaluation

Every action leaves evidence. Every behavior can be evaluated.

Models change. Prompts change. Tools change. Knowledge and policies change. So enterprise AI needs evaluation that keeps running, not a test at launch.

Always changingModelsPromptsToolsKnowledgePolicies
TaskTraceOutcome
Evaluation
  • CorrectnessIs the finding right?
  • Evidence qualityIs every claim sourced?
  • Policy complianceDid it stay inside the rules?
  • Tool behaviorWere tools called as expected?
  • CostWhat did the run cost?
  • LatencyHow long did it take?
Regression or improvementreviewed and released by people
Evaluation informs improvement. The system does not change itself silently; people review changes and release them.

Observability

Know what happened. And why.

Every step of an Investigation is recorded with who acted, in which role, with which capability, on which source and under which policy. It is the same principle as the rest of Deepin: evidence before decision.

INV-2026-0418 · Northwind Lojistik A.Ş.

Select a step to see the record behind it.

Human approval requested
Who acted
Investigation lead (agent)
Role
Investigation lead
Capability
Approval request
Source
Evidence package
Evidence
3 findings, 4 sources
Policy
Risk level changes require human approval
Decision
Waiting for the risk analyst
When
18 Sep 2026 · 09:42:24

Sample trace with a fictional company.

Harness and Spaces

One Harness. Domain-specific Spaces.

A Space packages reusable roles, capabilities, data sources, knowledge, policies, evaluations and Investigations for a business domain.

Deepin HarnessOrganizationReasoningKnowledgeGovernanceEvaluation
deepin.risk→
  1. Company Investigation
  2. Evidence + findings
  3. Human decision
deepin.finance→
  1. Chargeback Investigation
  2. Evidence + recommendation
  3. Human decision
HarnessThe reusable enterprise runtime.
SpaceThe implementation for one business domain.
InvestigationThe concrete business work.

In use

Built underneath real Investigations.

Harness capabilities are being shaped by concrete enterprise Investigations, not in an isolated framework experiment. Each reference is shown at the maturity it has actually reached.

deepin.riskCompany Investigation
  • Production
  • Production
  • Production
  • Embedded
deepin.financeChargeback Investigation
  • Contracted · In development
ContractedEmbedded or deployingIn production

R&D

Research-backed architecture.

Deepin Harness is continuously developed through R&D in organizational reasoning, governed agent collaboration, constraint-aware planning and continuous evaluation.

  1. 01

    Organizational Reasoning

    Reasoning according to role, authority, organizational state and constraints.

  2. 02

    Governed Collaboration

    Delegation and collaboration between roles while preserving authority boundaries.

  3. 03

    Constraint-Aware Planning

    Planning toward goals while respecting policies, authority and operational constraints.

  4. 04

    Continuous Evaluation

    Evaluating traces, outcomes and behavior as models, tools and organizational context change.

TÜBİTAK-supported R&D

Deepin Harness

Build AI that can work inside your organization.

Start from a real Investigation, or use Deepin Harness to bring governed AI into an existing enterprise workflow.

Bring us an Investigation Talk about Deepin Harness
info@deepin.space

See it in a Space: deepin.risk → deepin.finance →

Don't deploy another agent. Build an organization.

Deepin

Get to know Deepin technology.

Deepin Harness is part of Deepin. Deepin builds agents that investigate business cases, show the evidence and wait for a person to approve the decision.

See the other Spaces, the Harness underneath and how it all works on the Deepin home page.

Explore Deepin