For developers, it is very important to understand the evolution of agent building. This can be understood step-by-step through three distinct phases.
Step 1: The 2024-2025 Framework Era, where building an agent meant heavy lifting; developers relied on frameworks like LangChain, manually wrote tool wrappers, and stood up complex RAG pipelines to compensate for short foundation model context windows. During this time, 80% of a project involved writing middle-layer glue code instead of focusing on the agent’s actual job.
Step 2: The Capability Shift, driven by three major changes that collapsed this middle layer of development: model SDKs started shipping with built-in tools (like Bash, Grep, and WebSearch) out of the box, progressive “skills” replaced heavy startup tool registries to save context tokens, and RAG was demoted as larger context windows and direct file access made vector search a fallback rather than a default.
Step 3: The 2026 SDK-First Era, where the new default playbook is to start simple; developers now rely directly on model SDKs and MCP servers for core capabilities, reserving heavy frameworks strictly for complex edge cases like multi-provider routing, deterministic typing, or multi-agent orchestration.
Ultimately, the hard infrastructure problems became cheap and expected, moving the differentiator up a layer and allowing engineering teams to focus purely on the unique value their agents actually deliver.
I really liked this in-depth article by Pulumi on the three eras of agent building: <a href="https://www.pulumi.com/blog/how-building-ai-agents-has-changed/" target="_blank" rel="noopener nofollow">https://www.pulumi.com/blog/how-building-ai-agents-has-changed/</a>

Pavan Belagatti

A concise breakdown of how AI agent development has evolved through three phases: the 2024-2025 Framework Era dominated by LangChain, manual tool wrappers, and RAG pipelines; a Capability Shift where model SDKs gained built-in tools and larger context windows reduced RAG dependency; and the emerging 2026 SDK-First Era where developers start simple with model SDKs and MCP servers, reserving heavy frameworks only for complex edge cases. The net result is that infrastructure complexity has collapsed, shifting the engineering differentiator to the unique value agents actually deliver.

Understanding the Three Eras of Agent Building