From Context to Intent: A Project Manager’s Guide to Context, Agentic, Loop, Graph and Intent Engineering

By Kiran Viswanatha | AI Leadership & Responsible AI

 

Artificial intelligence is moving through an important transition.

A few years ago, project managers primarily asked:

“How can AI help me generate a status report, summarize requirements, or create a project plan?”

Today, the more important question is becoming:

“How do I design and govern AI systems that can understand the project, make decisions, execute work, learn from outcomes, and remain aligned with business intent?”

That shift has introduced a new vocabulary: Context Engineering, Agentic AI, Loop Engineering, Graph Engineering, and Intent Engineering.

These terms can sound like technical jargon. For project managers, however, they represent something much more practical: different control points for managing AI-enabled work.

The terminology is also still evolving. Context engineering is now widely discussed as the discipline of supplying an AI system with the right information, tools and state; loop engineering focuses on iterative execution and verification; graph engineering makes workflow topology explicit; and intent engineering is emerging around encoding organizational goals, constraints and success criteria for agents.

Understanding the difference can help project managers move from using AI as a productivity tool to managing AI as a project participant and operational system.

The simplest way to remember the difference

These are not competing approaches.

They are complementary.

A mature AI-enabled project may use all five.

1. Context Engineering — Give AI the Right Project Reality

The first problem with AI is often not intelligence.

It is context.

A project manager may know:

• the approved scope
• stakeholder expectations
• project constraints
• previous decisions
• risks
• dependencies
• organizational policies
• budget limitations
• regulatory requirements
• lessons learned

But an AI model does not automatically know all of that.

Context engineering is about deliberately selecting, structuring and delivering the information an AI system needs at the moment it needs it. This can include documents, project state, memory, tool outputs, policies, constraints and enterprise data.

Think of it this way:

Prompt engineering:
“Create a project status report.”

Context engineering:
“Here is the approved baseline, current schedule, RAID log, stakeholder priorities, latest sprint results, budget status, previous steering committee decisions and reporting template. Now create the status report.”

The second approach is much closer to how an experienced project manager actually works.

Real-world example

Imagine an AI assistant is asked:

“Are we on track for the product launch?”

Without context, it might look at the schedule and say yes.

With engineered context, it might discover:

• development is 95% complete;
• testing is only 65% complete;
• a regulatory approval is pending;
• the critical vendor is two weeks late;
• the launch date cannot move because of a contractual commitment.

The answer changes dramatically.

The model didn't necessarily become smarter. The context became better.

IBM describes this challenge in enterprise agent environments as a situation where a model may be technically capable but still make unsafe decisions because it lacks the right business state, approval status or policy context.

Why PMs should care

Project managers are essentially context managers already.

You spend much of your day deciding:

Who needs to know what, when, and based on which version of reality?

That is remarkably similar to context engineering.

2. Agentic AI — Move From Answering to Acting

Traditional generative AI primarily responds.

Agentic AI can plan, reason, use tools, execute actions, evaluate results and continue toward a goal.

For example:

Traditional AI

PM:

“Create a risk register.”

AI:

“Here is a risk register.”

Agentic AI

PM:

“Identify emerging risks for the ERP implementation and prepare them for Friday's steering committee.”

An agent might:

1. inspect the project schedule;
2. analyze open defects;
3. review the RAID log;
4. check vendor milestones;
5. identify emerging patterns;
6. classify risks;
7. calculate potential impact;
8. draft mitigation recommendations;
9. update the risk repository;
10. prepare a steering committee summary;
11. request human approval before publishing.

That is a fundamentally different operating model.

Agentic workflows can coordinate multiple agents, tools and people, with nodes representing actions, agents, tools or decision logic.

The important PM question becomes:

“What authority am I giving the agent?”

This is where Responsible AI becomes essential.

An AI agent might be allowed to:

• draft a communication;
• create a Jira ticket;
• recommend a schedule adjustment;

but perhaps not:

• approve a million-dollar change;
• terminate a vendor;
• change contractual commitments;
• release production software;
• communicate regulatory decisions externally.

The project manager becomes part of the AI control architecture.

3. Loop Engineering — Don't Let AI Stop at the First Answer

Human project managers rarely operate in one step.

We work through cycles:

Plan → Execute → Check → Correct → Repeat

AI agents increasingly need similar mechanisms.

Loop engineering focuses on designing iterative agent workflows where the system can act, observe results, evaluate them and continue or correct itself until defined completion criteria are satisfied. IBM describes loop engineering as designing these repeated agentic cycles with appropriate prompting, checking and guidance.

Consider an AI testing agent.

Weak implementation

Generate test cases.

Done.

Loop-engineered implementation

Generate → Execute → Evaluate → Fix → Retest → Validate → Stop

For example:

1. Generate test cases.
2. Execute tests.
3. Identify failures.
4. Determine whether failures are defects or environment issues.
5. Generate corrections.
6. Execute again.
7. Check regression impact.
8. Stop only when acceptance criteria are satisfied.
9. Escalate unresolved issues to a human.

This is much closer to actual project execution.

PM lesson

A project plan should not merely define:

“AI will perform testing.”

It should define:

• What is the success criterion?
• What constitutes failure?
• How many retries are allowed?
• When does AI escalate?
• What evidence is required?
• Who approves completion?

Those are loop governance questions.

4. Graph Engineering — Turn Complex Work Into a Visible System

Now imagine that your project contains:

• a business analyst agent;
• a development agent;
• a testing agent;
• a security agent;
• a compliance reviewer;
• a human product owner.

They cannot simply work randomly.

Their activities have dependencies.

For example:

Requirements → Architecture → Development → Testing → Security → Business Approval → Release

But there may also be branches:

Testing Failed → Development → Testing

or:

Security Failed → Security Remediation → Security Review

This is where graph engineering becomes useful.

Graph engineering makes the workflow topology explicit through nodes, edges, state transitions, branches, joins and human intervention points.

A PM can visualize it like a WBS + workflow + RACI + decision tree.

For example:

|

 

The important insight is that the graph does not replace the agent.

The agent performs work inside the graph.

A loop may also exist inside one node.

The graph controls where the work goes.

The loop controls how a particular activity improves or repeats.

5. Intent Engineering — The Most Important Question: Why?

This may be the most important concept for project managers.

An AI system can be technically correct and still produce the wrong business outcome.

Why?

Because execution is not the same as intent.

Consider a customer-service agent.

Its instruction might be:

“Resolve customer issues as quickly as possible.”

That sounds reasonable.

But an agent optimizing only for speed could:

• issue excessive refunds;
• bypass approval policies;
• provide unauthorized discounts;
• close cases prematurely.

The real organizational intent might be:

Resolve legitimate customer issues quickly while protecting customer trust, complying with policy, controlling financial exposure and escalating exceptions.

That is a much richer specification.

Emerging work on intent engineering frames it around encoding organizational goals, values and trade-off hierarchies into agent infrastructure.

For a project manager, intent should include:

Objective

What outcome are we trying to achieve?

Constraints

What must the agent never violate?

Priorities

If objectives conflict, which wins?

Success criteria

How do we know the agent succeeded?

Escalation criteria

When must a human intervene?

Authority

What can the agent actually change?

The Five Concepts Working Together

Let's put everything into one example.

Scenario: AI-Assisted ERP Implementation

Imagine an organization implementing a new ERP platform.

Context Engineering

The AI receives:

• approved requirements;
• project schedule;
• architecture;
• budget;
• RAID log;
• test results;
• vendor commitments;
• regulatory requirements;
• organizational policies.

Question:
What does AI need to know?

Agentic AI

AI agents can:

• analyze requirements;
• identify conflicts;
• create work items;
• summarize testing;
• investigate defects;
• recommend corrective actions;
• coordinate information gathering.

Question:
What can AI do?

Loop Engineering

For testing:

Generate → Test → Analyze → Fix → Retest → Validate

Question:
How does AI know when the work is good enough?

Graph Engineering

The project workflow defines:

Requirements → Design → Build → Test → Security → Approval → Release

with branches, dependencies and human gates.

Question:
How does work move through the system?

Intent Engineering

The AI is governed by:

“Deliver the ERP implementation on the approved timeline while protecting financial controls, regulatory compliance, data integrity and user readiness.”

Question:
Why is AI doing this, and what trade-offs should it make?

This is where AI becomes particularly interesting for project professionals.

AI does not eliminate project management concepts. It amplifies them.

A Real-World Project Management Example: Software Release

Consider a software release scheduled for Friday.

A traditional PM might coordinate:

• development;
• QA;
• security;
• DevOps;
• business;
• change management.

An AI-enabled operating model could look like this:

Context

AI knows:

• release scope;
• acceptance criteria;
• open defects;
• deployment checklist;
• production dependencies;
• change-management policy.

Agentic

Agents perform:

• test analysis;
• defect classification;
• deployment-readiness analysis;
• release-note generation.

Loop

If tests fail:

Analyze → Fix → Retest → Validate

Graph

The workflow determines:

QA → Security → Business Approval → Change Approval → Deployment

Intent

The governing objective is:

Release only when business value can be delivered without violating security, compliance, quality or operational readiness thresholds.

That is no longer simply an AI assistant.

It is an AI-enabled project operating model.

Why Project Managers Should Learn This Now

You don't need to become an AI engineer.

But you do need to understand the architecture because your role is moving closer to orchestrating humans, AI agents, data and automated workflows.

There are at least five reasons.

1. You will manage AI-enabled projects

AI will increasingly become part of the delivery team rather than simply a productivity application.

2. You will need to define AI accountability

If an agent makes a recommendation or executes an action:

Who owns the outcome?

That is fundamentally a project governance question.

3. You will need better requirements

Traditional requirements say:

“The system shall generate a report.”

AI requirements increasingly need to say:

“The agent shall generate the report using approved data sources, cite evidence, identify uncertainty, comply with defined policies and escalate material exceptions.”

That's a much richer requirement.

4. You will manage AI risks differently

Traditional risk:

Vendor delivery delayed.

AI-era risk:

Agent incorrectly interprets project status and automatically triggers an escalation.

The second requires runtime controls, not just a mitigation plan.

5. You will become an orchestrator of intelligence

The PM of the future may coordinate:

People + AI Agents + Data + Workflows + Governance

rather than simply:

People + Tasks + Schedule

That is a significant evolution of the profession.

A Practical Framework for PMs

When introducing AI into a project, ask these five questions:

CONTEXT

Does the AI have the right information?

AGENT

What decisions and actions can AI perform?

LOOP

How does AI verify, correct and improve its work?

GRAPH

How does work move between humans, agents and systems?

INTENT

What outcome are we optimizing for—and what must never be compromised?

If you can answer those five questions, you are already thinking beyond AI adoption and toward AI operating-model design.

The Responsible AI Dimension

There is an important caution here.

More autonomy does not automatically mean more value.

As AI moves from generating content toward taking actions, governance needs to move closer to runtime execution. Enterprise agent systems need appropriate access controls, policy enforcement, observability, approval states and human intervention mechanisms.

For project managers, this means incorporating questions such as:

• What data can the agent access?
• What actions can it perform?
• What decisions require human approval?
• Can we reconstruct why an action occurred?
• What happens when the agent is uncertain?
• What happens when the agent fails repeatedly?
• Can we stop the agent?
• Are its decisions aligned with the project's objectives?
• Can we measure value rather than simply activity?

Autonomy without governance is automation at risk.

From Project Manager to AI Orchestrator

The biggest change may not be technological.

It may be professional.

Project managers have traditionally orchestrated:

Scope + Schedule + Cost + Quality + Risk + People + Stakeholders

The AI-enabled project manager increasingly orchestrates:

Context + Intent + Agents + Loops + Graphs + People + Governance

This doesn't make project management less important.

It makes it more important.

When work becomes partially autonomous, someone must define the destination, establish the guardrails, design the workflow, monitor performance and remain accountable for the outcome.

That is very close to the heart of project management.

The Takeaway

Don't think of Context, Agentic, Loop, Graph and Intent Engineering as five fashionable AI buzzwords.

Think of them as five lenses for designing and managing AI-enabled work:

Context tells AI what it needs to know.
Agentic AI determines what it can do.
Loops determine how it learns from outcomes and knows when to continue or stop.
Graphs determine how work flows.
Intent determines why it is doing the work and what boundaries define success.

And for project managers, perhaps the most important principle is this:

The future PM won't simply manage projects where AI is used. The PM will increasingly manage projects where AI itself participates in the execution.

That requires a new competency: orchestrating intelligence responsibly.

The opportunity is not to replace project management with AI.