What is an AI coding agent?
A model-centered software system that can plan or choose actions, call tools such as terminals and browsers, inspect results and continue across multiple steps toward a programming goal.
∞Agents + The Future
AI Coding Agents:
History, Architecture & Future.
An agent does more than answer. It holds state, uses tools, takes actions, observes results and decides what to do next inside designed boundaries.
Direct answers
What makes software agentic?
What changed after code completion?
Code models moved from suggesting the next tokens toward modifying repositories, running tests, reading errors, using external tools and revising their own work.
What stays human?
Accountability, architecture, product intent, security boundaries, evaluation design and the decision to accept or reject consequential changes remain human governance problems even when implementation is delegated.
01 / A new loop
Prompt → act → observe → evaluate → act again.
Model
Language and multimodal models interpret goals, generate plans and produce candidate actions.
Tools
Browsers, terminals, APIs, databases, code runners and applications turn model outputs into real-world effects.
Memory
Working context, retrieval and durable state give the system continuity beyond a single exchange.
Evaluation
Tests, policies, critics and humans decide whether an action is acceptable or should be revised.
02 / The thesis
The future of coding may look less like typing a program and more like designing a constitution for systems that act.
Developers still define architecture, invariants, interfaces and failure boundaries. More implementation can become delegated, tested and revised by machines operating inside those boundaries.
The durable skill is not prompting alone. It is expressing intent precisely, building feedback, designing tools, observing behavior and knowing when a system should stop.