Project delivery is a continuous process.
Like an assembly line, work flows seamlessly: an 8:00 AM start follows a 5:00 PM finish the previous day, and Monday naturally continues where Friday ended.
However, not every activity can be interrupted.
You can’t start boiling water at 4:55 PM, stop at 5:00 PM, and expect to make tea at 8:00 AM the next morning. Some processes must start and finish within the same uninterrupted working period.
This leads to two important rules.

If a process cannot be interrupted, this constraint must be explicitly programmed in the project delivery model:

• Some tasks require “immediately” after logic allowing a successor start no later than defined period (or without any delay).
• Other activities should not be scheduled to start unless there is sufficient time for it to finish within the available working period.

Traditional CPM dependency types (FS, SS, FF and SF) cannot model this behaviour. They require additional project data modelling capabilities.
If these technological constraints exist but are not represented correctly, then both what-if analysis and Monte Carlo Simulation will produce unreliable results because the underlying delivery model does not reflect reality.

Artificial working periods make projects slower.
Although organising work into predefined blocks usually simplify planning and control, it often introduces hidden delays between predecessor and successor activities. These delays may not immediately affect the delivery date, but they accumulate and consume valuable project float. When uncertainty inevitably occurs, that lost float is no longer available to absorb disruptions.
Common examples include:

• Critical features planned to start at the next Agile sprint instead of commencing them immediately after enabling features are ready;
• a four-hour task being allocated an entire working day simply because it is easier to plan;
• features developed during an Agile sprint waiting for the next production release;
• fixed review and approval periods that delay work for administrative rather than technical reasons.
  These are not technological constraints. They are programmed delays.

Effective project delivery requires project processes to be planned around real-world constraints. This is the foundation of the Resource Critical Path Method (RCPM), which focuses on accurate project data modelling. Rather than simplifying reality through unrealistic assumptions, RCPM enables project teams to model and manage complexity, resulting in more reliable planning, forecasting, and decision-making.

The Resource Critical Path Method (RCPM) addresses both challenges.

An RCPM delivery model represents project processes as it is can be executed. It introduces uninterrupted (“no-break”) periods only where they are genuinely required by technology or the nature of the work, while eliminating artificial time blocks that add no value.

The result is an optimised delivery model that more accurately reflects reality, preserves valuable project float, improves the reliability of scenario analysis and Monte Carlo Simulation, and ultimately enables faster and more predictable project delivery.

Generative AI has made remarkable progress over the last few years. However, as we move from simple chat interfaces to autonomous AI agents, a new optimisation challenge emerges.

Like project managers balancing scope, time, and cost, AI agents must balance three competing objectives:

Quality of Outcome

How well the agent achieves the desired objective.

Time to Outcome

How quickly the objective is achieved.

Token Cost

How much computation and AI usage is consumed along the way.

Traditional chatbots generate a single response.
AI agents execute a process.

They may:

• Break a problem into smaller tasks
• Search for information
• Validate assumptions
• Run multiple reasoning cycles
• Use specialised tools
• Coordinate with other agents

As a result, the quality of the final outcome depends not only on model intelligence but also on how much work the agent performs.

A high-quality agent may:

• Search multiple sources
• Compare alternative solutions
• Verify facts
• Perform self-critique
• Run several iterations before presenting a recommendation

The outcome is usually better, but token consumption increases significantly.
A cheaper agent may produce an answer after a single reasoning pass, reducing costs but increasing the risk of errors, hallucinations, or missed opportunities.

Suppose an agent is helping a project manager recover a delayed project. The fastest approach is to suggest the first recovery strategy it finds.

A higher-quality approach may:

• Analyse the schedule
• Identify resource constraints
• Evaluate multiple recovery options
• Simulate outcomes
• Compare risks and benefits

The recommendation is likely to be better, but the decision takes longer.

An agent can often reduce delivery time by consuming more tokens.

For example:

• Running several specialist agents in parallel
• Performing multiple searches simultaneously
• Using larger reasoning models
• Conducting additional validation steps

The user receives an answer faster, but the computational cost increases

This is remarkably similar to project acceleration, where faster delivery often requires additional resources.

Today, most discussions focus on model intelligence.

However, as AI agents become more autonomous, the real challenge shifts toward managing the agent itself.

Questions become:

• How much validation is enough?
• How many reasoning cycles should be allowed?
• When should an agent stop searching and make a decision?
• How much token budget should be allocated to a task?
• Is the additional quality worth the additional cost and delay?

These are fundamentally management questions rather than technology questions.

Many people evaluate AI systems by model capability alone.

For AI agents, a more meaningful measure may be:
Outcome Quality per Token per Unit of Time
The best agent is not necessarily the one that produces the most sophisticated answer.

It is the one that achieves the required outcome with the optimal balance of:

• Quality of outcome
• Time to outcome
• Token cost

However, there is a broader question beyond the AI Agent Iron Triangle.
Organisations do not simply decide how to configure an AI agent. They decide whether deploying AI agents is the right investment in the first place. The benchmark is not another AI system. The benchmark is the alternative use of the available resources.

For every dollar invested in AI agents, an organisation could instead invest in:

• Training and developing existing staff
• Hiring additional specialists
• Improving processes and governance
• Better project management methods and tools
• Knowledge management and organisational learning

As a result, the key question becomes:

Is investment in AI agents the most effective way to improve organisational performance?

The answer is not always yes.

In many situations, the root cause of poor outcomes is not a lack of intelligence but a lack of capability, experience, methodology, leadership, or organisational maturity.

For example, a project organisation with weak planning practices may achieve greater benefits from developing planner competencies than from deploying AI scheduling agents. Likewise, a poorly governed organisation may gain more from improving decision-making processes than from automating them.

This creates a second optimisation problem:
Should the organisation invest in artificial intelligence or human intelligence?

Just as project managers must decide where to allocate limited resources, executives must decide whether the next dollar is better spent on technology or on people. In many situations, investment in people remains the rational choice. AI agents may accelerate existing capabilities, but they rarely replace capabilities that do not already exist.

In other words, before optimising the AI Agent Iron Triangle, organisations must answer a more fundamental question:

Is AI the best investment, or would developing human capability deliver a greater return?

This is remarkably similar to project management itself. Success is rarely achieved by buying better software alone. More often, it comes from finding the right balance between people, processes, methods, and technology. AI agents are simply the latest participant in that optimisation problem.