The abstraction boundary moved

July 29, 2026

AI coding agents broke the old boundary: people once understood systems, wrote code, and handed execution to deterministic tools.

AI coding agents broke the old boundary: people once understood systems, wrote code, and handed execution to deterministic tools. Now people state intent; agents explore systems, choose tools, write code, test, and revise.

Intent → Agent action → Verified change

The engineering stack, translated

The familiar layers remain; their immediate user is now often an LLM.

Software engineering	Agent engineering
Documentation	AGENTS.md, context files
Libraries / plugins	Skills, tools, MCP servers
IDE	Agent workspace
Shell commands	Tool calls
CI / tests	Agent evaluations and verifiers
Permissions	Sandboxes and approval systems
Package discovery	Skill registries / marketplaces
Microservices	Multi-agent systems
APIs	Structured function calling
Version control	Agent checkpoints / worktrees
Developer onboarding	Context injection
Operating system	Agent runtime / orchestrator

A restart, not a replacement

Teams are rediscovering the durable concerns of software engineering under a new constraint: the program is probabilistic, context-limited, and can choose its actions. In that sense, software engineering is being restarted from around 1970 — with LLMs as the immediate tool users.

Dependency management
Capability discovery
Access control
Reproducibility
State management
Observability
Modularity
Trust and verification