Knowledge Sources

Connect the tools you already use

The Knowledge Graph grows richer with every integration. Start with Git and CI/CD. Layer in cloud, telemetry, security, and cost data as your use cases expand.

First-party (Harness)

CI PipelinesCD & GitOpsFeature FlagsCloud CostSecurity TestingChaos EngineeringIDPArtifact Registry

Third-party integrations

GitHub / GitLabAWS / GCP / AzureKubernetesJiraPagerDutyDatadogGoogle Developer Connect

Google Cloud Technology Partner of the Year— DevOps 2026

Expanded collaboration with Google Cloud

Harness has integrated Google Cloud's Developer Connect into the Software Delivery Knowledge Graph, giving joint customers a continuously updated, relationship-aware view spanning both platforms.

Read the announcement

What this integration delivers

Deployment event logs from Google Cloud in the graph

Runtime data and application dependency information

Faster diagnosis tracing problems to source files

Enterprise-grade access controls across both platforms

Harness AI on Vertex AI — available on Google Cloud Marketplace

Architecture

From raw data to operational reasoning

A semantic layer sits between your tools and your AI agents, translating raw events into structured, relationship-aware context your agents can act on.

Data Sources.

Git, CI/CD, K8s, cloud billing, incidents, policies

Knowledge Graph.

Entities, relationships, canonical identities, real-time sync

Semantic Layer.

Governed queries, RBAC filtering, structured + unstructured retrieval

AI Agents.

Expert Agents (chat + MCP) and Worker Agents (pipeline steps)

Both Expert Agents (in chat and MCP) and Worker Agents (in pipelines) read from the same Knowledge Graph. Context compounds with every deployment cycle.

From AI-assisted to AI-operational

The difference between helping and operating

AI-assisted DevOps helps write code, generate pipelines, and summarize logs — useful, but shallow. AI-operational DevOps understands how software actually moves from commit to production: constraints, dependencies, governance, and all. That shift only happens when the platform understands itself. The Knowledge Graph is what makes that possible.

Pipeline generation. AI generates pipelines aligned to org standards and governance rules, not guesswork.

Root-cause analysis. Traverses the full dependency chain automatically — no manual correlation across tools.

Safe rollbacks. Validated against the downstream dependency graph before executing.

Cost anomaly tracing. Spend spikes traced to the specific deployment decision that caused them.

AI-assisted (without Knowledge Graph)

💬 “Why did my deployment fail?”

Generic AI

Pages a senior engineer. 20 minutes of investigation across 4 tools.

Ask a question...

AI-operational (with Knowledge Graph)

💬 “Why did my deployment fail?”

Harness AI

“Pipeline #4471 failed because service-checkout lost access to the QA cluster at 14:32. Policy update iam-policy-update-032 revoked it. Affected: 2 downstream services. Fix: reinstate access in Terraform module eks-rbac.”

Ask Harness AI...

Identity & Normalization

One entity. Many names. One truth.

The same service is called something different in Git, Kubernetes, CloudWatch, and your runbook. Without normalization, AI agents can't connect the dots. The Knowledge Graph defines canonical service identities with rule-based matching and alias support — multiple teams keep their conventions without cluttering the graph.

Rule-based entity matching. Across first-party Harness modules and third-party systems.

Alias support. Teams keep their own naming conventions; the graph handles cross-system correlation.

Drift prevention. Change management keeps entity definitions consistent as services and namings evolve.

Use cases

What the Knowledge Graph makes possible

Why did it fail — and what should you do about it

When a pipeline fails, the Knowledge Graph lets agents traverse the dependency chain from build artifact → deployment → environment → policy → access control — answering not just what failed but why, in seconds. No paging an expert. No correlating across 4 tools.

Pipeline #4471 fails — agent queries the graph for execution state and connected entities

Identifies change delta — new artifact version, policy update, or access revocation since last success

Maps blast radius — which downstream services depend on this pipeline's output

Proposes fix — targeted remediation with the Terraform module or RBAC config to update

Knowledge Graph + RAG

Why Knowledge Graph + RAG beats RAG alone

RAG is powerful for unstructured document retrieval. Knowledge graphs add semantic structure and multi-hop reasoning. Together, they deliver what neither can alone.

RAG only

Text similarity without system understanding

Retrieves documents based on semantic similarity — no awareness of how systems relate

Cannot traverse dependency chains across entities (service → cluster → policy)

Stale document chunks — no real-time sync from live delivery systems

No governance — retrieval doesn't enforce RBAC or policy scope

Strong for: searching runbooks, docs, log summaries, natural language queries

✓ Recommended — Knowledge Graph + RAG

Structured reasoning + unstructured breadth

Graph queries traverse multi-hop relationships: pipeline → service → environment → owner

Near real-time sync from Harness and third-party systems. Context reflects the current state.

RBAC and OPA policies applied at the semantic layer — agents only see what they're permitted to see

Combines graph-structured retrieval with RAG for unstructured content attached to graph nodes

Compounding intelligence — each deployment cycle enriches the graph for future agent queries

Common pitfalls

Where Knowledge Graph initiatives fail

Three failure modes repeat across organizations. Avoiding them is the difference between a graph that ships value in 30 days and one that becomes technical debt.

Overmodeling

Modeling everything before solving anything

Teams model 100 entities before picking a use case. The graph becomes academic and unused. Start with 10 entities that answer one real question.

Fix: Pick 1–2 use cases first. Model only what's required.

Undermodeling

Missing the relationships that create value

Skipping key relationships — like linking a service to its owning team or the policy that governs its deployment — means AI gives shallow, incorrect answers.

Fix: Map the relationships for your chosen use case before expanding.

Stale Context

Perfectly modeled data, a week old

During an outage, teams need to know why it's failing now, not what the state was last month. Stale data is as dangerous as no data.

Fix: Near real-time sync is non-negotiable for delivery workflows.

AI is only as powerful as the context behind it