2025 Retrospective: The Year of Agents

2025 is ending. It was the year AI moved from “impressive demos” to “daily tools.” Here’s what happened and what it means.

The Defining Trend: Agents

What Changed

2024: “ChatGPT can write code” 2025: “AI agents complete tasks”

Before: Write prompt → Get code → Copy-paste → Test → Debug → Repeat
After: Describe goal → Agent plans → Agent executes → Review result

The loop closed. AI went from a suggestion tool to a task executor.

Key Agent Developments

Reasoning Models Arrived

o1, R1, and others that “think”:

# 2024 approach
prompt = "Solve this: [complex math problem]"
# Often wrong on multi-step

# 2025 approach
prompt = "Solve this: [complex math problem]"
# Model thinks first, gets it right more often

The unlock: Extended inference compute improves accuracy.

Local AI Became Real

What Changed

• Models run on phones (Phi-3, Gemma)
• 7B models on laptops (R1 distilled)
• Apple Intelligence (on-device)
• Edge deployment viable

Why It Matters

• Privacy by default
• Zero latency
• Offline capability
• Cost-effective at scale

Developer Tools Transformed

Daily Workflow Changes

Productivity Impact

Honest assessment:

• Routine coding: 2-3x faster
• Novel problems: Modest improvement
• System design: Still human-led
• Debugging production: AI helps but expertise needed

Open Source Caught Up

Notable Models

The Implication

OpenAI no longer has a monopoly on frontier capabilities.

The Regulation Era Began

EU AI Act

• Prohibited practices: February 2025
• Transparency: August 2025
• High-risk compliance: August 2026

Global Response

• China: AI governance rules
• US: Executive order, state laws
• UK: “Pro-innovation” approach
• Others: Watching and adapting

Industry Dynamics

Big Tech AI

Startup Landscape

• Harder to compete on model quality
• Differentiation through application
• Vertical AI companies thriving
• Infrastructure plays maturing

What Didn’t Happen

Overhyped Predictions That Missed

Why

Demos ≠ Production. Scaling labs ≠ reliable deployment.

Personal Learning Stack

What I Used Daily

1. Claude for complex reasoning
2. Cursor for coding
3. Perplexity for research
4. Local Llama for private queries
5. Gemini for multimodal

What Changed My Work

• Less typing, more reviewing
• Better first drafts
• Faster learning new tech
• More time on design, less on implementation

Technical Highlights

Favorite Announcements

1. DeepSeek R1: Open reasoning
2. Django 5.2 LTS: Composite PKs
3. Python 3.14: Deferred annotations
4. uv: Fast Python packaging
5. Gateway API GA: K8s networking

Underrated Developments

• Structured outputs (better JSON)
• Voice mode improvements
• Memory and personalization
• Tool use reliability

Looking Forward to 2026

Predictions

What I’m Watching

• Test-time compute scaling
• Multi-agent collaboration
• AI safety research
• Open source vs closed dynamics

Final Thoughts

2025 was transformative but not revolutionary. AI became a daily tool—useful, sometimes frustrating, still requiring human judgment.

The year of agents means: AI can do more, we supervise more, and the value is in what we choose to build.

Build something. The tools are ready.

The future arrived gradually, then suddenly. Then gradually again.