For the past few years, our interaction with artificial intelligence has been defined by the prompt. We ask a question, the AI answers. Generative AI is undeniably powerful, but it stays fundamentally reactive — it waits for instructions. That creates a hidden tax on productivity: manual coordination. Whether you are planning a complex multi-city trip or hiring a specialised engineer, you are still the one breaking down tasks, checking results and initiating every subsequent step.
That overhead is real cognitive load. Agentic AI matters because it shifts the paradigm from a system that answers questions to one that autonomously achieves goals. If you want the ground-level definitions first, start with our complete guide to agentic AI. This piece goes a layer deeper: how the machinery actually works, and why it changes your job rather than just your tooling.
1. Proactivity vs. reactivity: the “go to Goa” lesson
The core distinction between traditional GenAI and agentic AI is initiative. Planning a trip to Goa with a chatbot, the AI is a passive participant. You ask for flight options, then hotels, then itineraries. You are the operator, manually driving the process step by step.
An agentic system is proactive. You give it a single goal — “organise a five-day trip to Goa for three people within a $2,000 budget” — and the agent takes over. It plans, takes action, adapts to changes, and seeks help only when necessary. That moves you from operator to supervisor: the system manages the complexity of the workflow while you own the strategy.
2. The candidate plan: why an agent needs a search strategy
Sophisticated agents treat planning as a search problem. Given a goal such as filling a backend engineer vacancy, the agent does not follow a linear script. It treats the journey from the initial state (the vacancy) to the final state (the hire) as a path-optimisation challenge, in three steps.
Generation
The agent creates multiple candidate plans. Plan A might post on job portals like LinkedIn; Plan B might prioritise internal referrals or specialised hiring agencies.
Evaluation
Each plan is scored against concrete metrics:
- Efficiency — which path reaches the final state fastest?
- Tool availability — does the agent actually have the APIs it needs (LinkedIn API access, for example) to execute this plan?
- Cost and risk — which plan fits the budget and carries the lowest chance of failure?
Selection
Based on that evaluation, the agent commits to the most optimised path. The important detail is that the plan is chosen, not hardcoded — which is exactly what lets it be re-chosen later when reality disagrees.
3. Reasoning is the glue between planning and execution
If planning is the map, reasoning is the cognitive glue that lets the agent navigate it. Reasoning is a continuous loop: the agent interacts with its environment — the job market, candidate responses, digital tools — and processes feedback through a repeating cycle of interpret information → draw conclusions → make decisions.
That loop is what enables three things:
- Goal decomposition — breaking a high-level goal into actionable sub-tasks.
- Tool selection — reasoning about which tool fits this moment: a résumé parser to screen candidates, versus a calendar API to schedule the interview. Giving agents reliable access to those tools is its own problem, which is what the Model Context Protocol sets out to solve.
- Adaptability — where the agent shows its real power. If the environment returns negative feedback, such as a week passing with zero applicants, it does not stall. It might propose a mid-course goal change: switch the search from a full-time employee to a freelancer so the project stays on track.
4. Controlled autonomy: the human-in-the-loop safeguard
A common misconception is that more autonomy means less control. The opposite is true: highly autonomous systems demand more deliberate management. In an agentic architecture the human is the supervisor — the final approver of high-risk actions.
Unbounded autonomy is genuinely dangerous. An agent could burn a large ad budget or send an incorrect offer letter. Autonomy is therefore managed through four pillars:
- Permissions and scope — defining exactly which tools the agent may use independently.
- Human-in-the-loop (HITL) — mandatory checkpoints where the agent must pause for approval, such as before rolling out a final offer.
- Guardrails — hard ethical or organisational rules, e.g. never schedule interviews on weekends.
- Override controls — the human keeps ultimate power through pause and stop, able to halt or alter behaviour at any moment.
5. Short-term vs. long-term memory
To function across a multi-week project, an agent needs context awareness through a dual-memory system.
- Short-term memory holds session-specific data and immediate tool output — “the résumé parser indicates Candidate B has 3 years of Django experience.”
- Long-term memory holds persistent goals, organisational policy and user preferences that apply across every session — “the company always prefers remote candidates” or “never send an offer letter without human approval.”
6. The five-part anatomy of an agent
A working mental model of any agentic system comes down to five components:
| Component | Role |
|---|---|
| Brain (LLM) | Goal interpretation, reasoning loops and natural-language communication. |
| Orchestrator (workflow) | The nervous system: sequences tasks, handles conditional routing (if step A fails, do step B) and manages retries. |
| Tools (APIs / RAG) | The hands and legs — calling APIs and retrieving domain-specific data. Grounding an agent in your own data is exactly what a RAG pipeline provides. |
| Memory (storage) | Tracks state, progress and historical context so the agent does not forget the goal. |
| Supervisor (human interface) | Facilitates HITL interactions, enforces guardrails and handles escalation on edge cases. |
7. From prompters to managers
The shift to agentic AI is a genuine evolution in how software gets used. We are moving from being prompters who supply granular, repetitive instructions to being managers who define high-level goals and oversee autonomous workflows. The value is in reducing cognitive load: the system handles the how so you can concentrate on the why. It is already visible in practice — see how AI is taking on cloud operations toil, or how to hand an entire workflow to Claude.
Which leaves the question worth sitting with: if an AI could autonomously handle 80% of your most complex project today, what would you do with the 20% of your time that is left?