DEV Community

bayu priatno
bayu priatno

Posted on

The Missing Infrastructure for AI Engineering: Knowledge

Why AI Agents Need More Than Context Windows

The software engineering industry is experiencing a fundamental shift.

AI coding agents are no longer limited to autocomplete.

They can understand repositories, modify multiple files, execute commands, run tests, investigate failures, refactor systems, generate documentation, and increasingly coordinate complex engineering tasks.

This creates enormous leverage.

But there is a problem that becomes more visible as AI agents become more capable:

An intelligent agent is only as effective as the engineering knowledge available to it.

This is why the next challenge in AI-native software engineering may not be model intelligence.

It may be knowledge infrastructure.

And this is one of the foundational ideas behind NAEOS — Nusantara AI Engineering Operating System.


1. The Context Problem

When developers work on a mature software system, they rarely rely only on the source code.

They also understand things such as:

  • why a particular architecture was chosen
  • which components are considered critical
  • which APIs must remain backward compatible
  • which dependencies are prohibited
  • how authentication works
  • which security controls are mandatory
  • why a particular workaround exists
  • how deployments are performed
  • which incidents happened in the past
  • which technical decisions were already made
  • which business rules constrain the implementation

Much of this knowledge is not encoded directly in the codebase.

It exists across:

Architecture Documents
       +
ADRs
       +
Documentation
       +
Tickets
       +
Pull Requests
       +
Runbooks
       +
Policies
       +
Chat Discussions
       +
Incident Reports
       +
Developer Experience
       +
Organizational Memory
Enter fullscreen mode Exit fullscreen mode

A senior engineer may connect all of these mentally.

An AI agent generally cannot.

And this creates a major engineering gap.


2. Code Is Not the Entire System

A common assumption in AI-assisted development is:

"If the agent can access the repository, it understands the project."

That assumption is dangerous.

A repository tells an agent what the system currently does.

It does not necessarily tell the agent:

why it does it.

Consider a simple example.

An AI agent discovers:

timeout = 30
Enter fullscreen mode Exit fullscreen mode

The agent might reasonably decide that 30 seconds is too high and reduce it to 10.

From a code perspective, this may look like an improvement.

But perhaps the value exists because:

  • an external payment provider requires it
  • a legacy integration has a known latency profile
  • a previous production incident established the constraint
  • a downstream system has a strict retry sequence
  • a business workflow depends on the timeout

The code contains the value.

The engineering knowledge contains the reason.

This distinction is critical.

Code represents implementation. Knowledge represents intent, constraints, decisions, and relationships.

AI engineering needs both.


3. The Knowledge Fragmentation Problem

Most organizations already have enormous amounts of engineering knowledge.

The problem is that it is fragmented.

A typical enterprise might have:

GitHub
├── Source Code
├── Pull Requests
├── Issues
└── Discussions

Wiki
├── Architecture
├── Standards
└── Documentation

Ticketing System
├── Requirements
├── Bugs
└── Decisions

Cloud Platform
├── Infrastructure
├── Configuration
└── Logs

CI/CD
├── Pipelines
├── Tests
└── Deployment History

Chat
├── Decisions
├── Troubleshooting
└── Tribal Knowledge
Enter fullscreen mode Exit fullscreen mode

Each system contains part of the organizational memory.

But these systems are rarely treated as a unified knowledge system.

The result is:

Knowledge exists
        ↓
Knowledge is fragmented
        ↓
Knowledge becomes difficult to retrieve
        ↓
Humans reconstruct context
        ↓
AI receives incomplete context
        ↓
AI makes weaker decisions
Enter fullscreen mode Exit fullscreen mode

This is one of the problems NAEOS is designed to address.


4. From Documents to Knowledge Objects

Traditional documentation treats knowledge primarily as documents.

AI-native engineering requires a more structured approach.

Instead of thinking only in terms of:

Document
Enter fullscreen mode Exit fullscreen mode

we can think in terms of:

Knowledge Object
Enter fullscreen mode Exit fullscreen mode

A Knowledge Object has an identity, type, metadata, lifecycle, ownership, version, relationships, validation status, and content.

For example:

id: knowledge.security.authentication

type: security-policy

version: 2.1.0

status: approved

owner:
  team: security-engineering

scope:
  organization: naeos
  projects:
    - platform
    - api

content:
  principle: >
    All externally exposed APIs must use approved authentication mechanisms.

constraints:
  - mTLS where required
  - OAuth2 for delegated authorization
  - no plaintext credentials

relations:
  - architecture.api-gateway
  - standard.api-security
  - blueprint.service
Enter fullscreen mode Exit fullscreen mode

This changes the role of knowledge.

It becomes something that can be:

  • indexed
  • searched
  • validated
  • versioned
  • linked
  • governed
  • retrieved by AI
  • audited
  • reused

Knowledge becomes infrastructure.


5. Why Metadata Matters

A paragraph of text is not enough for enterprise AI engineering.

An AI system needs to know:

What is this?

Who owns it?

Is it approved?

Which version is active?

When was it updated?

What scope does it apply to?

What does it depend on?

What depends on it?

Can an AI agent use it?

Can it be modified?

What policies govern it?
Enter fullscreen mode Exit fullscreen mode

Therefore, metadata becomes a first-class component of the knowledge architecture.

A mature Knowledge Object may contain:

Identity
Ownership
Domain
Classification
Version
Status
Scope
Priority
Dependencies
References
Relationships
Permissions
Validation
Lifecycle
Audit
Integrity
Enter fullscreen mode Exit fullscreen mode

This makes knowledge machine-readable and governable.


6. Knowledge Has a Lifecycle

Not every piece of information should immediately become authoritative.

Engineering knowledge evolves.

A useful lifecycle is:

Draft
  ↓
Validation
  ↓
Review
  ↓
Approval
  ↓
Publish
  ↓
Active
  ↓
Deprecated
  ↓
Archived
Enter fullscreen mode Exit fullscreen mode

This matters because AI agents should distinguish between:

Approved Architecture
Enter fullscreen mode Exit fullscreen mode

and:

Someone's Old Architecture Proposal
Enter fullscreen mode Exit fullscreen mode

Without lifecycle metadata, an AI system may retrieve both and treat them as equally authoritative.

That is unacceptable in production engineering.

Therefore:

Knowledge retrieval must consider authority, not only semantic similarity.


7. Knowledge Versioning

Engineering knowledge changes.

An architecture decision from two years ago may no longer apply.

A security policy may be updated.

An API contract may evolve.

A deployment strategy may change.

Therefore, knowledge must be versioned.

For example:

architecture.service-boundary

v1.0.0
    ↓
v1.1.0
    ↓
v1.2.0
    ↓
v2.0.0
Enter fullscreen mode Exit fullscreen mode

Versioning enables:

  • history
  • diff
  • rollback
  • traceability
  • compatibility analysis
  • reproducibility
  • auditability

It also gives AI agents something extremely important:

temporal context.

The agent can understand not only:

"What is the rule?"

but also:

"Which rule was valid when this implementation was created?"


8. Knowledge Is Relational

Another important limitation of document-based systems is that engineering knowledge does not exist independently.

Architecture affects APIs.

APIs affect tests.

Security policies affect infrastructure.

Blueprints affect implementations.

Incidents affect policies.

Policies constrain workflows.

Workflows invoke agents.

This creates relationships.

For example:

Security Policy
      │
      │ constrains
      ▼
API Architecture
      │
      │ implemented by
      ▼
Service Blueprint
      │
      │ generates
      ▼
Repository
      │
      │ validated by
      ▼
Security Tests
Enter fullscreen mode Exit fullscreen mode

This is a graph.

And this graph is extremely valuable for AI agents.


9. The Knowledge Graph

NAEOS therefore treats the Knowledge Graph as an important architectural component.

A Knowledge Graph can represent:

Nodes:
- Architecture
- Policy
- API
- Service
- Blueprint
- Plugin
- Workflow
- Incident
- Decision
- Standard
- Repository
- Agent
- Deployment

Edges:
- depends_on
- constrains
- implements
- references
- supersedes
- generated_by
- validated_by
- owned_by
- deployed_to
- affected_by
Enter fullscreen mode Exit fullscreen mode

A simplified graph:

                ┌───────────────┐
                │   Policy      │
                └───────┬───────┘
                        │
                     constrains
                        │
                        ▼
                ┌───────────────┐
                │ Architecture  │
                └───────┬───────┘
                        │
                    defines
                        │
                        ▼
                ┌───────────────┐
                │   Blueprint   │
                └───────┬───────┘
                        │
                   generates
                        │
                        ▼
                ┌───────────────┐
                │   Repository  │
                └───────┬───────┘
                        │
                   validated by
                        │
                        ▼
                ┌───────────────┐
                │     Tests     │
                └───────────────┘
Enter fullscreen mode Exit fullscreen mode

Now an AI agent can perform more than semantic search.

It can perform relationship-aware retrieval.


10. Semantic Search Is Not Enough

Modern RAG systems often work like this:

Question
   ↓
Embedding
   ↓
Vector Search
   ↓
Top-K Documents
   ↓
LLM
Enter fullscreen mode Exit fullscreen mode

This is useful.

But enterprise engineering requires more.

Suppose an engineer asks:

"How should authentication be implemented for the payment service?"

A purely semantic system might retrieve documents containing:

  • authentication
  • OAuth
  • payment
  • security

But the correct answer may require connecting:

Payment Service
       ↓
Service Architecture
       ↓
Security Policy
       ↓
Authentication Standard
       ↓
API Contract
       ↓
Approved Implementation Pattern
Enter fullscreen mode Exit fullscreen mode

This is why an AI-native knowledge system should support multiple retrieval strategies.


11. Hybrid Knowledge Retrieval

A mature retrieval engine can combine:

Semantic Search
        +
Keyword Search
        +
Metadata Search
        +
Graph Search
        +
Vector Search
        +
Relationship Traversal
        +
Authority Ranking
        +
Recency
Enter fullscreen mode Exit fullscreen mode

Conceptually:

                User Intent
                     │
                     ▼
             Query Understanding
                     │
       ┌─────────────┼─────────────┐
       ▼             ▼             ▼
   Semantic       Metadata       Graph
    Search         Search        Search
       │             │             │
       └─────────────┼─────────────┘
                     ▼
               Result Fusion
                     │
                     ▼
              Relevance Ranking
                     │
                     ▼
              Context Assembly
                     │
                     ▼
                  AI Agent
Enter fullscreen mode Exit fullscreen mode

The objective is not to retrieve the most similar documents.

The objective is to retrieve the most relevant engineering context.


12. Knowledge and AI Memory Are Different

Another important distinction is between knowledge and memory.

They are related, but they should not be treated as identical.

Knowledge

Represents validated and reusable engineering information.

Examples:

Architecture Standards
Security Policies
API Contracts
Engineering Constitution
Approved Runbooks
Enter fullscreen mode Exit fullscreen mode

Memory

Represents contextual experience and historical information.

Examples:

Previous Agent Actions
Recent Tasks
Conversation Context
Past Failures
Execution History
Temporary Context
Enter fullscreen mode Exit fullscreen mode

A useful conceptual model is:

                 AI SYSTEM
                     │
          ┌──────────┴──────────┐
          │                     │
      Knowledge               Memory
          │                     │
   Validated Truth        Context & Experience
          │                     │
          └──────────┬──────────┘
                     ▼
              Context Engine
                     │
                     ▼
                 AI Agent
Enter fullscreen mode Exit fullscreen mode

This separation becomes important for governance.

Not every memory should automatically become organizational knowledge.


13. The Knowledge Promotion Pipeline

One of the interesting possibilities is allowing useful experience to become validated knowledge.

Imagine an AI agent encounters a recurring production problem.

It discovers:

Problem
  ↓
Investigation
  ↓
Solution
  ↓
Validation
  ↓
Review
Enter fullscreen mode Exit fullscreen mode

The resulting lesson could become a Knowledge Object:

Incident
   ↓
Root Cause
   ↓
Engineering Lesson
   ↓
Candidate Knowledge
   ↓
Human Review
   ↓
Approved Knowledge
Enter fullscreen mode Exit fullscreen mode

This creates a learning loop.

Engineering Activity
        ↓
Observation
        ↓
Experience
        ↓
Candidate Knowledge
        ↓
Validation
        ↓
Organizational Knowledge
        ↓
Future AI Context
Enter fullscreen mode Exit fullscreen mode

This is how an engineering organization can begin to compound knowledge.


14. From RAG to Engineering Context

RAG is often described as:

Retrieve documents and give them to the model.

But production AI engineering requires a richer concept.

The system should construct an Engineering Context.

For example:

Engineering Context
│
├── Current Task
├── Repository
├── Architecture
├── Applicable Policies
├── Relevant Standards
├── Dependencies
├── Previous Decisions
├── Security Constraints
├── Testing Requirements
├── Deployment Constraints
├── Historical Incidents
└── Agent Memory
Enter fullscreen mode Exit fullscreen mode

Then the AI agent operates against that context.

This produces a much stronger architecture:

Intent
  +
Engineering Context
  +
Policy
  +
Tools
  +
Memory
  ↓
AI Agent
  ↓
Engineering Action
Enter fullscreen mode Exit fullscreen mode

15. Knowledge Governance

Enterprise knowledge must also be governed.

Consider the following questions:

Who can create knowledge?

Who can approve it?

Who can modify it?

Who can deprecate it?

Who can access sensitive knowledge?

Which AI agents can consume it?

Which projects can use it?

What happens when a policy conflicts with another policy?
Enter fullscreen mode Exit fullscreen mode

These are not merely documentation questions.

They are governance questions.

Therefore, the Knowledge System should integrate:

RBAC
ABAC
Tenant Isolation
Access Policies
Classification
Encryption
Audit Trail
Digital Signatures
Integrity Checks
Approval Workflow
Compliance
Enter fullscreen mode Exit fullscreen mode

Knowledge must be treated with the same seriousness as other enterprise assets.


16. Security of AI Knowledge

This becomes especially important when AI agents can access organizational knowledge.

Not all knowledge should be exposed equally.

For example:

Public
Internal
Confidential
Restricted
Highly Restricted
Enter fullscreen mode Exit fullscreen mode

An agent working on a public API may need:

API Standards
Architecture Guidelines
Documentation Standards
Enter fullscreen mode Exit fullscreen mode

But it may not need:

Production Secrets
Security Incident Details
Customer Data
Internal Credentials
Restricted Infrastructure Information
Enter fullscreen mode Exit fullscreen mode

Therefore, retrieval itself must be security-aware.

A simplified model:

AI Agent
   ↓
Identity
   ↓
Authorization
   ↓
Knowledge Policy
   ↓
Allowed Knowledge Scope
   ↓
Retrieval
Enter fullscreen mode Exit fullscreen mode

The principle is straightforward:

An AI agent should retrieve only the knowledge it is authorized to access.


17. Knowledge Quality

A knowledge system can become dangerous if it simply accumulates information.

More knowledge does not automatically mean better knowledge.

The system must evaluate:

Validity
Accuracy
Authority
Freshness
Consistency
Completeness
Duplication
References
Ownership
Security Classification
Enter fullscreen mode Exit fullscreen mode

A Knowledge Validator can therefore perform:

Schema Validation
      ↓
Reference Validation
      ↓
Policy Validation
      ↓
Consistency Validation
      ↓
Duplicate Detection
      ↓
Semantic Validation
      ↓
Approval
Enter fullscreen mode Exit fullscreen mode

This creates an important principle:

Knowledge must be engineered, not merely stored.


18. The Knowledge Compiler

An interesting architectural concept follows from this.

If knowledge is structured, validated, versioned, and related, it can become executable engineering context.

This suggests a Knowledge Compiler.

Conceptually:

Raw Knowledge
      ↓
Parse
      ↓
Normalize
      ↓
Validate
      ↓
Resolve References
      ↓
Build Relationships
      ↓
Compile
      ↓
Engineering Context
Enter fullscreen mode Exit fullscreen mode

The compiled output could be consumed by:

  • AI agents
  • workflow engines
  • policy engines
  • code generators
  • validators
  • documentation systems
  • runtime components

This creates a bridge between static engineering knowledge and executable engineering workflows.


19. Knowledge as a Control Mechanism

Once knowledge becomes structured and connected to policies, it can influence execution.

Consider:

Architecture
      ↓
Policy
      ↓
Workflow
      ↓
Agent
      ↓
Execution
Enter fullscreen mode Exit fullscreen mode

The agent does not simply receive instructions.

It receives a governed environment.

For example:

Task:
Implement a new API endpoint.

Applicable Knowledge:
- API Standard v3
- Security Policy v5
- Architecture Decision ADR-042
- Testing Standard v2

Required Gates:
- Unit tests
- Integration tests
- Security validation
- API contract validation
- Documentation

Allowed Tools:
- Repository
- Test Runner
- Static Analyzer

Restricted:
- Production credentials
- Direct production deployment
Enter fullscreen mode Exit fullscreen mode

This is much closer to how an experienced engineering organization operates.


20. The Role of NAEOS

This is where NAEOS enters the picture.

NAEOS is being designed around the idea that AI engineering needs a shared system containing:

Governance
Constitution
Architecture
Knowledge
Policies
Memory
Workflows
Blueprints
Quality Gates
Agent Instructions
Plugins
Observability
Audit
Versioning
Enter fullscreen mode Exit fullscreen mode

These components create an engineering environment around AI agents.

The simplified model is:

                    NAEOS
                      │
       ┌──────────────┼──────────────┐
       │              │              │
   Knowledge       Policies       Memory
       │              │              │
       └──────────────┼──────────────┘
                      ▼
               Context Engine
                      │
                      ▼
                AI Agents
                      │
                      ▼
               Engineering
                      │
                      ▼
               Validation
                      │
                      ▼
                 Feedback
                      │
                      ▼
                  Knowledge
Enter fullscreen mode Exit fullscreen mode

This is the foundation of an AI-native engineering operating model.


21. Why This Is Different From a Wiki

At first glance, someone might ask:

"Isn't this just a better documentation platform?"

No.

A documentation platform primarily helps humans read information.

A Knowledge System for AI engineering must help machines reason and operate within engineering constraints.

The difference can be summarized as:

Traditional Documentation AI Engineering Knowledge
Human-readable Human + machine-readable
Document-centric Knowledge-object-centric
Mostly static Versioned and lifecycle-driven
Search-oriented Retrieval + relationship-oriented
Passive Operational
Limited relationships Knowledge graph
Manual context assembly Automated context assembly
Documentation Engineering infrastructure

This is a fundamental architectural difference.


22. The Future: Knowledge-Native Engineering

If AI agents become increasingly autonomous, engineering organizations may eventually operate around knowledge systems in much the same way modern organizations operate around source-control systems and cloud infrastructure.

The stack could look something like:

┌───────────────────────────────────────┐
│             HUMAN INTENT             │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│              GOVERNANCE               │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│             KNOWLEDGE                 │
│ Architecture · Policy · Standards    │
│ Domain · Decisions · Memory          │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│          ENGINEERING CONTROL          │
│ Planning · Validation · Quality      │
│ Workflow · Security · Permissions    │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│             AI AGENTS                 │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│          ENGINEERING SYSTEMS          │
│ Code · APIs · Infrastructure · Tests │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│         OBSERVABILITY & AUDIT         │
└───────────────────┬───────────────────┘
                    ↓
┌───────────────────────────────────────┐
│        KNOWLEDGE FEEDBACK LOOP        │
└───────────────────┴───────────────────┘
Enter fullscreen mode Exit fullscreen mode

The engineering organization becomes a continuously learning system.


23. The Bigger Idea

The important question is not:

"How do we give AI more context?"

It is:

"How do we build an organizational knowledge infrastructure that AI can safely and reliably operate against?"

That is a much bigger question.

It involves:

  • information architecture
  • knowledge graphs
  • metadata
  • governance
  • access control
  • semantic retrieval
  • vector search
  • versioning
  • validation
  • memory
  • policy engines
  • workflow orchestration
  • AI agents
  • observability
  • auditability

In other words:

Knowledge becomes part of the engineering platform.


24. Where NAEOS Is Going

NAEOS is an exploration of this direction.

The goal is not to build another chatbot.

It is not to build another autocomplete engine.

It is not to replace every existing AI coding agent.

Instead, the objective is to establish an engineering layer where different AI agents can operate using a common:

Knowledge.

Policy.

Architecture.

Memory.

Workflow.

Governance.

Quality model.

The agent can change.

The model can change.

The vendor can change.

The engineering system remains.

That separation could become increasingly important in an AI-native software industry.


Conclusion

The first phase of AI-assisted development was about making AI better at writing code.

The next phase is about making AI better at engineering systems.

That requires more than larger context windows.

It requires structured organizational knowledge.

It requires knowledge that can be:

identified, versioned, validated, linked, governed, retrieved, and continuously improved.

Because ultimately:

AI does not just need more context. AI needs the right engineering context.

And engineering context should not live inside someone's memory, a forgotten document, an old Slack message, or an isolated prompt.

It should become infrastructure.

That is one of the foundational ideas behind NAEOS.

Build the knowledge layer.
Govern the agents.
Standardize the engineering.
Let the system learn.

The future of AI-native software engineering may depend not only on how intelligent our models become, but on how intelligently we build the systems around them.

Top comments (0)