If 2024 was the year we learned to prompt, 2025 was the year the work itself changed shape.
The through-line
Look back across the year and a single arc dominates: the unit of AI-assisted work moved from the suggestion to the task. We started 2025 accepting autocompletions one line at a time. We ended it delegating whole features to agents that plan, edit, run, and verify on their own. The word the industry settled on for doing this well — agentic engineering — captures the shift from ad-hoc prompting to building systems that direct AI deliberately.
The moments that mattered
A few markers from the year worth remembering:
- Agents grew up. Coding assistants went from autocomplete to autonomous loops that handle multi-step work, with enterprises reporting large feature work completed in hours rather than weeks.
- The frontier sped up. By late 2025 a major model was landing roughly every week — Grok, Gemini, Claude, GPT all leapfrogging — turning frontier capability into a fast-moving commodity.
- MCP became infrastructure. The Model Context Protocol went from one vendor's idea to mainstream agent plumbing, and by December was donated to a Linux Foundation fund — shared infrastructure, not a product.
- The toolchain caught up. React shipped a compiler that automates memoization; frameworks leaned into server-first rendering; the boring-but-crucial ergonomics improved across the board.
What we learned the hard way
The year also delivered a dose of realism:
- The harness matters more than the model. Most agent failures were configuration failures, not reasoning failures.
- There is a verification tax. Studies found experienced developers could actually go slower with AI on some tasks, paying for it in review and debugging. Generation is solved; correctness is not.
- Vibes don't scale. Tossing ad-hoc prompts at a model is cheap to start and expensive to maintain. Structure — specs, evals, guardrails — is what holds up in production.
The defining lesson of 2025: AI makes the first 80% nearly free, and makes the last 20% the entire job.
Where it leaves us
The role did not disappear; it moved up the stack. The most valuable engineers ended the year spending less time typing implementations and more time on judgment, architecture, verification, and designing the systems agents work inside. The craft did not die — it relocated.
Looking ahead
The open questions heading into 2026 are not about whether agents can generate code. They can. They are about trust, verification, security, and economics: how we know agent output is correct, how we keep the agent surface safe, and how we make all of it pay. Those are engineering problems — which is exactly why this is still our job.
Here is to a year that changed the work, and to doing the next one with more structure and less hype.