Area 4

Agentic coding
& AI agents

The most visceral shift of 2026: tools that plan, act, observe, and iterate until the job is done.

Definition

An agent is an LLM with tools and a loop.

An agent is a model placed inside infrastructure that gives it tools, memory, and an action loop. A long system prompt alone does not make an agent — the loop and tool access do.

🧰

Tools

It can call functions, run code, query systems.

🧠

Memory

It carries context across steps within a task.

🔁

A loop

It repeats until the goal is met — not a one-shot reply.

The agent loop

Perceive → plan → act → observe.

The canonical cycle, rooted in the Russell & Norvig agent definition (1995) and made practical for LLMs by the ReAct framework (Yao et al., 2022), which showed +34% on ALFWorld from interleaving reasoning and action. Typically one loop iteration = one LLM call + one tool call.

1

Perceive

Read the current state and goal.

2

Plan

Reason about the next action.

3

Act

Call a tool or run code.

4

Observe

Read the result, then repeat.

Human-in-the-loop vs autonomous

Scale autonomy with risk.

A spectrum from approve-every-step to fully autonomous. Autonomy level should scale with task risk and test coverage.

LOWER RISK / MORE CONTROL

Approve every step

You confirm each action. Right for unfamiliar code, destructive operations, or thin test coverage.

HIGHER THROUGHPUT

Fully autonomous

Multi-hour tasks spawning sub-agents — e.g. Nubank reportedly used a fleet of Devins to migrate ~6M lines of code.

Background & parallel agents

Throughput's frontier.

Cloud agents fix a GitHub issue in a sandbox while you're in meetings; IDEs run multiple agents in parallel (Antigravity, Cursor 2.0 up to 8).

Multi-agent orchestration

Supervisor + specialists.

A supervisor delegates to planner / executor / verifier agents. More capable on complex workflows — but it multiplies token cost and observability needs.

Token consumption vs a normal chat

Agentic and multi-agent workloads consume far more tokens
Normal chat
Single agent
~4×
Multi-agent
up to ~15×